queue.json

{
  "bridge": [
    {
      "arxiv_id": "2601.03067",
      "title": "Joint Encoding of KV-Cache Blocks for Scalable LLM Serving",
      "abstract": "Modern large language models (LLMs) drive interactive AI systems but are bottlenecked by the memory-heavy growth of key-value (KV) caches, which limits real-time throughput under concurrent loads. Existing KV-cache compression methods rely on rigid heuristics, disrupt tensor layouts, or require specialized compute, hindering scalability and deployment.   We propose joint encoding of KV-cache blocks, which fuses similar blocks across requests and input chunks into shared representations while preserving standard cache structure. This alleviates the KV-cache memory bottleneck, supporting high-concurrency serving without specialized hardware. Theoretically, we analyze the rate-distortion tradeoff of fused cache blocks under a Poisson process model. Empirically, our method achieves up to 4.38 $\\times$ KV-cache compression with negligible accuracy loss across diverse LLMs and benchmarks, outperforming recent structured and adaptive compression baselines. In real LLM serving, joint encoding improves the token throughput by $\\sim$40\\% on a single-machine vLLM benchmark, demonstrating substantial gains in inference throughput. Code is available at https://github.com/sef1/kv_fast_fusion  kv_joint_encoding.",
      "authors": [
        "Joseph Kampeas",
        "Emir Haleva"
      ],
      "published_at": "2026-01-06T14:50:58",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03067v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5128501057624817,
      "sim_memory": 0.6588089466094971,
      "sim_negative": 0.16072729229927063,
      "broad_relevance": 0.5128501057624817,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.37791067361831665,
      "evidence_score": 0.5793147180559945,
      "direct_memory_relevance": 0.6588089466094971,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.77,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5744370306727911,
      "memory_score": 0.6459604501495684,
      "quality_score": 0.5727258872223978,
      "bridge_score": 0.5744370306727911,
      "max_axis_score": 0.6459604501495684,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Theory",
        "Inference"
      ],
      "status": "Cover now",
      "why_now": "KV-cache growth is one of the most immediate bottlenecks in high-concurrency LLM serving, and this paper proposes a deployable compression approach that preserves standard cache structure while showing throughput gains in vLLM. It squarely hits current production concerns around memory capacity, serving efficiency, and system scalability without requiring exotic hardware.",
      "why_not_higher": "The topic is more infrastructure-focused than broadly model-architectural, so it may not land with every general AI listener. The empirical story is strong but still early, with limited evidence yet on very large-scale production heterogeneity, long-context extremes, and operational tradeoffs beyond the reported benchmarks.",
      "downgrade_reasons": [
        "serving-systems topic is narrower than frontier model capability papers",
        "evidence appears centered on single-machine vLLM evaluation",
        "compression/fusion methods in KV cache are an active crowded area"
      ],
      "what_would_raise_priority": "Independent replication across more serving stacks, larger models, multi-node deployments, and long-context workloads would make this an even stronger must-cover paper.",
      "one_sentence_episode_hook": "What if LLM servers could stop treating every request's KV cache as isolated baggage and instead fuse similar cache blocks for a 4.38x memory win and roughly 40% more throughput?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03043",
      "title": "Lil: Less is Less When Applying Post-Training Sparse-Attention Algorithms in Long-Decode Stage",
      "abstract": "Large language models (LLMs) demonstrate strong capabilities across a wide range of complex tasks and are increasingly deployed at scale, placing significant demands on inference efficiency. Prior work typically decomposes inference into prefill and decode stages, with the decode stage dominating total latency. To reduce time and memory complexity in the decode stage, a line of work introduces sparse-attention algorithms. In this paper, we show, both empirically and theoretically, that sparse attention can paradoxically increase end-to-end complexity: information loss often induces significantly longer sequences, a phenomenon we term ``Less is Less'' (Lil). To mitigate the Lil problem, we propose an early-stopping algorithm that detects the threshold where information loss exceeds information gain during sparse decoding. Our early-stopping algorithm reduces token consumption by up to 90% with a marginal accuracy degradation of less than 2% across reasoning-intensive benchmarks.",
      "authors": [
        "Junhao Hu",
        "Fangze Li",
        "Mingtao Xu",
        "Feifan Meng",
        "Shiju Zhao",
        "Tiancheng Hu",
        "Ting Peng",
        "Anmin Liu",
        "Wenrui Huang",
        "Chenxu Liu"
      ],
      "published_at": "2026-01-06T14:23:58",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03043v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 2,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5221384167671204,
      "sim_memory": 0.5686707496643066,
      "sim_negative": 0.24759384989738464,
      "broad_relevance": 0.5221384167671204,
      "momentum": 0.031333575283049045,
      "teachability": 0.85,
      "novelty_score": 0.3494328260421753,
      "evidence_score": 0.529861228866811,
      "direct_memory_relevance": 0.5686707496643066,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5418023483230938,
      "memory_score": 0.4247845723633093,
      "quality_score": 0.5889444915467243,
      "bridge_score": 0.4247845723633093,
      "max_axis_score": 0.5418023483230938,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Adjacent",
        "Systems",
        "Inference",
        "Theory"
      ],
      "status": "Monitor",
      "why_now": "Sparse-attention decode tricks are actively being explored to cut long-context serving cost, and this paper argues a counterintuitive failure mode: making attention cheaper can lengthen generations enough to erase the win. That is timely for anyone evaluating post-training inference optimizations on reasoning-heavy LLM workloads.",
      "why_not_higher": "The memory/storage link is real but indirect: the core contribution is an evaluation and stopping-policy warning about sparse decode, not a new cache/offload/paging mechanism. It also looks more like a benchmark-plus-mitigation result than a broadly adopted serving recipe at this stage.",
      "downgrade_reasons": [
        "memory relevance is adjacent rather than core",
        "focuses on sparse-attention decode rather than a general serving stack change",
        "unclear breadth across production models and hardware settings",
        "may be a cautionary benchmark result more than a lasting systems primitive"
      ],
      "what_would_raise_priority": "Show robust gains across major open and closed model families, realistic long-context serving traces, and direct end-to-end latency/memory improvements versus standard KV-cache-efficient baselines.",
      "one_sentence_episode_hook": "What if making attention cheaper actually makes your LLM slower by causing it to talk longer?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03229",
      "title": "SpANNS: Optimizing Approximate Nearest Neighbor Search for Sparse Vectors Using Near Memory Processing",
      "abstract": "Approximate Nearest Neighbor Search (ANNS) is a fundamental operation in vector databases, enabling efficient similarity search in high-dimensional spaces. While dense ANNS has been optimized using specialized hardware accelerators, sparse ANNS remains limited by CPU-based implementations, hindering scalability. This limitation is increasingly critical as hybrid retrieval systems, combining sparse and dense embeddings, become standard in Information Retrieval (IR) pipelines. We propose SpANNS, a near-memory processing architecture for sparse ANNS. SpANNS combines a hybrid inverted index with efficient query management and runtime optimizations. The architecture is built on a CXL Type-2 near-memory platform, where a specialized controller manages query parsing and cluster filtering, while compute-enabled DIMMs perform index traversal and distance computations close to the data. It achieves 15.2x to 21.6x faster execution over the state-of-the-art CPU baselines, offering scalable and efficient solutions for sparse vector search.",
      "authors": [
        "Tianqi Zhang",
        "Flavio Ponzina",
        "Tajana Rosing"
      ],
      "published_at": "2026-01-06T18:15:53",
      "categories": [
        "cs.DB",
        "cs.AR"
      ],
      "url": "http://arxiv.org/abs/2601.03229v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "other",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.3850385546684265,
      "sim_memory": 0.3980174660682678,
      "sim_negative": 0.20490583777427673,
      "broad_relevance": 0.3850385546684265,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.4463194012641907,
      "evidence_score": 0.5493147180559945,
      "direct_memory_relevance": 0.3980174660682678,
      "systems_leverage": 0.5499999999999999,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.5599999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.34635487449145563,
      "memory_score": 0.6497230059871997,
      "quality_score": 0.5727258872223978,
      "bridge_score": 0.34635487449145563,
      "max_axis_score": 0.6497230059871997,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Bridge",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Hybrid retrieval pipelines are making sparse search relevant again, and this paper tackles it with a concrete near-memory architecture on CXL-style hardware where data movement is the core bottleneck. It fits the current industry interest in memory-centric AI systems beyond just dense model serving.",
      "why_not_higher": "The work is more about vector database infrastructure than mainstream model training or LLM serving, so broad AI audience pull is limited. Its impact also depends on specialized near-memory hardware adoption rather than an immediately transferable software technique.",
      "downgrade_reasons": [
        "narrowly focused on sparse ANNS",
        "hardware-specific deployment path",
        "indirect relevance to mainstream transformer workflows",
        "appears primarily as a systems benchmark paper"
      ],
      "what_would_raise_priority": "Stronger evidence that the ideas transfer to production hybrid retrieval stacks or general-purpose memory-bound AI serving systems would raise it.",
      "one_sentence_episode_hook": "If sparse retrieval is back, do we need to move vector search off the CPU and into memory itself?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03324",
      "title": "Bare-Metal Tensor Virtualization: Overcoming the Memory Wall in Edge-AI Inference on ARM64",
      "abstract": "The deployment of Large Language Models (LLMs) on edge devices is fundamentally constrained by the \"Memory Wall\" the bottleneck where data movement latency outstrips arithmetic throughput. Standard inference runtimes often incur significant overhead through high-level abstractions, dynamic dispatch, and unaligned memory access patterns. In this work, we present a novel \"Virtual Tensor Core\" architecture implemented in software, optimized specifically for ARM64 microarchitectures (Apple Silicon). By bypassing standard library containers in favor of direct memory mapping (mmap) and implementing hand-tuned NEON SIMD kernels, we achieve a form of \"Software-Defined Direct Memory Access (DMA).\" Our proposed Tensor Virtualization Layout (TVL) guarantees 100% cache line utilization for weight matrices, while our zero-copy loader eliminates initialization latency. Experimental results on a 110M parameter model demonstrate a stable throughput of >60 tokens/second on M2 hardware. While proprietary hardware accelerators (e.g., Apple AMX) can achieve higher peak throughput, our architecture provides a fully open, portable, and deterministic reference implementation for studying the memory bottleneck on general-purpose ARM silicon, meeting the 200ms psycholinguistic latency threshold without opaque dependencies.",
      "authors": [
        "Bugra Kilictas",
        "Faruk Alpay"
      ],
      "published_at": "2026-01-06T15:00:40",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.AR",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03324v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.3978739082813263,
      "sim_memory": 0.559910237789154,
      "sim_negative": 0.1955457329750061,
      "broad_relevance": 0.3978739082813263,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.48185181617736816,
      "evidence_score": 0.26,
      "direct_memory_relevance": 0.559910237789154,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.4,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.34163994491100313,
      "memory_score": 0.5979730713367462,
      "quality_score": 0.575,
      "bridge_score": 0.34163994491100313,
      "max_axis_score": 0.5979730713367462,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Memory/Storage Core",
        "Systems",
        "Hardware",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Edge LLM inference on commodity ARM64 hardware is timely, and the paper squarely targets the memory-wall problem with concrete claims about cache-line utilization, zero-copy loading, and bandwidth-aware tensor layout. It could make a useful episode if the implementation is real and reproducible because it offers an open alternative to opaque vendor accelerators.",
      "why_not_higher": "The evidence appears narrow: a single 110M-parameter model on Apple M2 with ambitious systems claims but limited validation against broader workloads, baselines, and hardware. The framing is highly Apple/ARM-specific and some terminology ('software-defined DMA', 'virtual tensor core') sounds more rhetorical than established.",
      "downgrade_reasons": [
        "single-model evaluation",
        "limited hardware diversity",
        "unclear baseline rigor",
        "Apple/ARM64 specificity",
        "strong marketing-style terminology",
        "uncertain novelty versus existing mmap/zero-copy/SIMD runtime work"
      ],
      "what_would_raise_priority": "A stronger cross-model, cross-hardware evaluation with clear ablations against standard runtimes and quantified memory-bandwidth/cache effects would raise this substantially.",
      "one_sentence_episode_hook": "Can careful tensor layout, mmap, and hand-tuned NEON kernels let a plain ARM CPU punch above its weight for LLM inference by attacking the memory wall directly?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03199",
      "title": "DIP: Dynamic In-Context Planner For Diffusion Language Models",
      "abstract": "Diffusion language models (DLMs) have shown strong potential for general natural language tasks with in-context examples. However, due to the bidirectional attention mechanism, DLMs incur substantial computational cost as context length increases. This work addresses this issue with a key discovery: unlike the sequential generation in autoregressive language models (ARLMs), the diffusion generation paradigm in DLMs allows \\textit{efficient dynamic adjustment of the context} during generation. Building on this insight, we propose \\textbf{D}ynamic \\textbf{I}n-Context \\textbf{P}lanner (DIP), a context-optimization method that dynamically selects and inserts in-context examples during generation, rather than providing all examples in the prompt upfront. Results show DIP maintains generation quality while achieving up to 12.9$\\times$ inference speedup over standard inference and 1.17$\\times$ over KV cache-enhanced inference.",
      "authors": [
        "Yang Li",
        "Han Meng",
        "Chenan Wang",
        "Haipeng Chen"
      ],
      "published_at": "2026-01-06T17:24:16",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03199v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5605237483978271,
      "sim_memory": 0.5769320726394653,
      "sim_negative": 0.3419671654701233,
      "broad_relevance": 0.5605237483978271,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.4599868059158325,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.5769320726394653,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.82,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.547155145406723,
      "memory_score": 0.30907962179183956,
      "quality_score": 0.595,
      "bridge_score": 0.30907962179183956,
      "max_axis_score": 0.547155145406723,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Adjacent",
        "Systems",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Diffusion language models are still emerging, and this paper offers a practical serving-side idea: dynamically changing which in-context examples are present during generation to cut inference cost. That makes it timely as people explore non-autoregressive alternatives and prompt-time efficiency tricks.",
      "why_not_higher": "The memory/storage angle is indirect: this is mainly context management for DLM inference, not a concrete KV-cache, paging, offload, or bandwidth systems paper. Its impact also depends on diffusion language models becoming more mainstream, and the abstract alone does not establish broad realistic deployment evidence.",
      "downgrade_reasons": [
        "memory connection is adjacent rather than core",
        "depends on adoption of diffusion language models",
        "unclear realism of workloads and hardware evaluation from abstract",
        "speedup claim may be architecture-specific"
      ],
      "what_would_raise_priority": "Show strong results on widely used DLMs and realistic serving setups, with clear comparisons to standard prompt compression, retrieval, and cache-aware inference baselines.",
      "one_sentence_episode_hook": "What if diffusion language models could rewrite their own prompt on the fly, keeping quality while making long-context inference dramatically cheaper?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03204",
      "title": "InfiAgent: An Infinite-Horizon Framework for General-Purpose Autonomous Agents",
      "abstract": "LLM agents can reason and use tools, but they often break down on long-horizon tasks due to unbounded context growth and accumulated errors. Common remedies such as context compression or retrieval-augmented prompting introduce trade-offs between information fidelity and reasoning stability. We present InfiAgent, a general-purpose framework that keeps the agent's reasoning context strictly bounded regardless of task duration by externalizing persistent state into a file-centric state abstraction. At each step, the agent reconstructs context from a workspace state snapshot plus a fixed window of recent actions. Experiments on DeepResearch and an 80-paper literature review task show that, without task-specific fine-tuning, InfiAgent with a 20B open-source model is competitive with larger proprietary systems and maintains substantially higher long-horizon coverage than context-centric baselines. These results support explicit state externalization as a practical foundation for stable long-horizon agents. Github Repo:https://github.com/ChenglinPoly/infiAgent",
      "authors": [
        "Chenglin Yu",
        "Yuchen Wang",
        "Songmiao Wang",
        "Hongxia Yang",
        "Ming Li"
      ],
      "published_at": "2026-01-06T17:35:57",
      "categories": [
        "cs.AI",
        "cs.MA"
      ],
      "url": "http://arxiv.org/abs/2601.03204v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5043788552284241,
      "sim_memory": 0.3724733591079712,
      "sim_negative": 0.22019587457180023,
      "broad_relevance": 0.5043788552284241,
      "momentum": 0.017490950964499406,
      "teachability": 0.75,
      "novelty_score": 0.5200152099132538,
      "evidence_score": 0.4893147180559946,
      "direct_memory_relevance": 0.3724733591079712,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.9199999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5932113359568144,
      "memory_score": 0.2830597738991107,
      "quality_score": 0.6027258872223977,
      "bridge_score": 0.2830597738991107,
      "max_axis_score": 0.5932113359568144,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Adjacent",
        "Bridge",
        "Systems",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Long-horizon agents are a live topic, and this paper offers a practical alternative to ever-growing context windows by externalizing persistent state into files. That design is timely because many teams are hitting reliability and cost limits in agent workflows built around pure prompt history.",
      "why_not_higher": "The memory angle is architectural rather than a direct KV-cache, paging, or bandwidth systems result, so it is adjacent rather than core to the podcast's storage lens. Evidence is promising but still limited to agent benchmarks and literature-review style workloads rather than broad real-world deployments.",
      "downgrade_reasons": [
        "Memory/storage link is indirect",
        "Evaluation breadth is limited",
        "Agent framework space is crowded",
        "Not a direct serving or hardware result"
      ],
      "what_would_raise_priority": "Stronger evidence across diverse real-world agent tasks plus clear comparisons on cost, latency, failure recovery, and context-length scaling would raise it to cover-now territory.",
      "one_sentence_episode_hook": "What if the way to make agents run for days isn't a bigger context window at all, but treating state like a filesystem and rebuilding the mind from snapshots?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03417",
      "title": "Implicit Graph, Explicit Retrieval: Towards Efficient and Interpretable Long-horizon Memory for Large Language Models",
      "abstract": "Long-horizon applications increasingly require large language models (LLMs) to answer queries when relevant evidence is sparse and dispersed across very long contexts. Existing memory systems largely follow two paradigms: explicit structured memories offer interpretability but often become brittle under long-context overload, while latent memory mechanisms are efficient and stable yet difficult to inspect. We propose LatentGraphMem, a memory framework that combines implicit graph memory with explicit subgraph retrieval. LatentGraphMem stores a graph-structured memory in latent space for stability and efficiency, and exposes a task-specific subgraph retrieval interface that returns a compact symbolic subgraph under a fixed budget for downstream reasoning and human inspection. During training, an explicit graph view is materialized to interface with a frozen reasoner for question-answering supervision. At inference time, retrieval is performed in latent space and only the retrieved subgraph is externalized. Experiments on long-horizon benchmarks across multiple model scales show that LatentGraphMem consistently outperforms representative explicit-graph and latent-memory baselines, while enabling parameter-efficient adaptation and flexible scaling to larger reasoners without introducing large symbolic artifacts.",
      "authors": [
        "Xin Zhang",
        "Kailai Yang",
        "Hao Li",
        "Chenyue Li",
        "Qiyu Wei",
        "Sophia Ananiadou"
      ],
      "published_at": "2026-01-06T21:10:10",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03417v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4132891595363617,
      "sim_memory": 0.4329555630683899,
      "sim_negative": 0.25163236260414124,
      "broad_relevance": 0.4132891595363617,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5272353291511536,
      "evidence_score": 0.36000000000000004,
      "direct_memory_relevance": 0.4329555630683899,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.82,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.49307204723358156,
      "memory_score": 0.17288666892051696,
      "quality_score": 0.595,
      "bridge_score": 0.17288666892051696,
      "max_axis_score": 0.49307204723358156,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Adjacent",
        "Bridge",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Long-context LLMs still struggle when relevant evidence is sparse and scattered, and this paper offers an interpretable retrieval-memory hybrid that speaks to a live design question in agentic and long-horizon systems. The latent-plus-explicit graph split is a potentially useful framing for balancing efficiency, stability, and inspectability.",
      "why_not_higher": "The memory angle is mostly architectural and retrieval-centric rather than a concrete systems contribution around KV cache, offload, paging, or bandwidth. Evidence appears benchmark-driven on long-horizon QA rather than showing clear impact on realistic serving stacks or production-scale deployments.",
      "downgrade_reasons": [
        "memory connection is adjacent, not core systems memory",
        "appears centered on benchmarked long-horizon QA",
        "limited evidence of hardware/serving implications",
        "crowded area of long-context and retrieval-memory hybrids"
      ],
      "what_would_raise_priority": "A stronger case that this method changes real LLM serving or agent memory design at scale, especially with concrete efficiency and latency wins on realistic deployments, would raise priority.",
      "one_sentence_episode_hook": "Can LLMs keep long-horizon memories in a latent graph yet reveal only the tiny symbolic subgraph you need\u2014getting both efficiency and interpretability without drowning in context?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03017",
      "title": "MMFormalizer: Multimodal Autoformalization in the Wild",
      "abstract": "Autoformalization, which translates natural language mathematics into formal statements to enable machine reasoning, faces fundamental challenges in the wild due to the multimodal nature of the physical world, where physics requires inferring hidden constraints (e.g., mass or energy) from visual elements. To address this, we propose MMFormalizer, which extends autoformalization beyond text by integrating adaptive grounding with entities from real-world mathematical and physical domains. MMFormalizer recursively constructs formal propositions from perceptually grounded primitives through recursive grounding and axiom composition, with adaptive recursive termination ensuring that every abstraction is supported by visual evidence and anchored in dimensional or axiomatic grounding. We evaluate MMFormalizer on a new benchmark, PhyX-AF, comprising 115 curated samples from MathVerse, PhyX, Synthetic Geometry, and Analytic Geometry, covering diverse multimodal autoformalization tasks. Results show that frontier models such as GPT-5 and Gemini-3-Pro achieve the highest compile and semantic accuracy, with GPT-5 excelling in physical reasoning, while geometry remains the most challenging domain. Overall, MMFormalizer provides a scalable framework for unified multimodal autoformalization, bridging perception and formal reasoning. To the best of our knowledge, this is the first multimodal autoformalization method capable of handling classical mechanics (derived from the Hamiltonian), as well as relativity, quantum mechanics, and thermodynamics. More details are available on our project page: MMFormalizer.github.io",
      "authors": [
        "Jing Xiong",
        "Qi Han",
        "Yunta Hsieh",
        "Hui Shen",
        "Huajian Xin",
        "Chaofan Tao",
        "Chenyang Zhao",
        "Hengyuan Zhang",
        "Taiqiang Wu",
        "Zhen Zhang"
      ],
      "published_at": "2026-01-06T13:42:51",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03017v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.44381266832351685,
      "sim_memory": 0.29899269342422485,
      "sim_negative": 0.3236953020095825,
      "broad_relevance": 0.44381266832351685,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6132620275020599,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.29899269342422485,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.515133104622364,
      "memory_score": 0.005697808027267437,
      "quality_score": 0.595,
      "bridge_score": 0.005697808027267437,
      "max_axis_score": 0.515133104622364,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Bridge"
      ],
      "status": "Monitor",
      "why_now": "Multimodal reasoning is a live topic, and this paper pushes autoformalization beyond text into diagrams and physics scenes, which is a genuinely interesting direction for AI audiences. It also introduces a benchmark that could help expose where frontier multimodal models still fail at grounding formal reasoning.",
      "why_not_higher": "The contribution is mostly a task framework and benchmark for a specialized reasoning setting, not something that obviously changes how mainstream AI systems are built or served. Evidence is promising but small-scale, with 115 curated samples and evaluation centered on frontier model performance rather than broad, reproducible systems validation.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Primarily a narrow multimodal formalization task",
        "Benchmark size is limited",
        "Unclear downstream impact on mainstream model building or serving"
      ],
      "what_would_raise_priority": "A larger benchmark plus evidence that the formalization pipeline materially improves real theorem proving, scientific assistants, or multimodal agent reliability would raise its priority.",
      "one_sentence_episode_hook": "Can an AI look at a physics or geometry diagram and turn it into machine-checkable mathematics, not just plausible text?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "public": [
    {
      "arxiv_id": "2601.03027",
      "title": "Reducing Hallucinations in LLMs via Factuality-Aware Preference Learning",
      "abstract": "Preference alignment methods such as RLHF and Direct Preference Optimization (DPO) improve instruction following, but they can also reinforce hallucinations when preference judgments reward fluency and confidence over factual correctness. We introduce F-DPO (Factuality-aware Direct Preference Optimization), a simple extension of DPO that uses only binary factuality labels. F-DPO (i) applies a label-flipping transformation that corrects misordered preference pairs so the chosen response is never less factual than the rejected one, and (ii) adds a factuality-aware margin that emphasizes pairs with clear correctness differences, while reducing to standard DPO when both responses share the same factuality. We construct factuality-aware preference data by augmenting DPO pairs with binary factuality indicators and synthetic hallucinated variants. Across seven open-weight LLMs (1B-14B), F-DPO consistently improves factuality and reduces hallucination rates relative to both base models and standard DPO. On Qwen3-8B, F-DPO reduces hallucination rates by five times (from 0.424 to 0.084) while improving factuality scores by 50 percent (from 5.26 to 7.90). F-DPO also generalizes to out-of-distribution benchmarks: on TruthfulQA, Qwen2.5-14B achieves plus 17 percent MC1 accuracy (0.500 to 0.585) and plus 49 percent MC2 accuracy (0.357 to 0.531). F-DPO requires no auxiliary reward model, token-level annotations, or multi-stage training.",
      "authors": [
        "Sindhuja Chaduvula",
        "Ahmed Y. Radwan",
        "Azib Farooq",
        "Yani Ioannou",
        "Shaina Raza"
      ],
      "published_at": "2026-01-06T14:01:34",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03027v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 2,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.3803938627243042,
      "sim_memory": 0.28131216764450073,
      "sim_negative": 0.21838413178920746,
      "broad_relevance": 0.3803938627243042,
      "momentum": 0.031333575283049045,
      "teachability": 0.75,
      "novelty_score": 0.5752854645252228,
      "evidence_score": 0.6198612288668109,
      "direct_memory_relevance": 0.28131216764450073,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.94,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.6231568778827061,
      "memory_score": 0.013576997757367537,
      "quality_score": 0.6189444915467244,
      "bridge_score": 0.013576997757367537,
      "max_axis_score": 0.6231568778827061,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training"
      ],
      "status": "Cover now",
      "why_now": "Hallucination reduction is a live concern for anyone shipping LLMs, and this proposes a simple drop-in modification to DPO rather than a heavyweight new alignment stack. The reported gains across multiple open-weight models and on out-of-distribution truthfulness benchmarks make it timely for practitioners evaluating post-training recipes.",
      "why_not_higher": "This is not a memory/storage paper, and much of the contribution is an incremental alignment objective plus dataset relabeling rather than a broader systems shift. The evidence is strong for factuality benchmarks, but the paper does not yet show downstream product impact or tradeoffs on helpfulness, style, and preference quality in real deployments.",
      "downgrade_reasons": [
        "No direct memory or storage mechanism",
        "Primarily an alignment/post-training method",
        "Benchmark-heavy with limited deployment evidence",
        "Unclear tradeoffs versus general preference optimization objectives"
      ],
      "what_would_raise_priority": "Independent replication on widely used post-training stacks plus clearer results on real assistant workloads and robustness-helpfulness tradeoffs would raise it further.",
      "one_sentence_episode_hook": "What if standard preference tuning is teaching LLMs to sound right instead of be right\u2014and a tiny DPO fix can cut hallucinations 5x?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03190",
      "title": "Maximizing Local Entropy Where It Matters: Prefix-Aware Localized LLM Unlearning",
      "abstract": "Machine unlearning aims to forget sensitive knowledge from Large Language Models (LLMs) while maintaining general utility. However, existing approaches typically treat all tokens in a response indiscriminately and enforce uncertainty over the entire vocabulary. This global treatment results in unnecessary utility degradation and extends optimization to content-agnostic regions. To address these limitations, we propose PALU (Prefix-Aware Localized Unlearning), a framework driven by a local entropy maximization objective across both temporal and vocabulary dimensions. PALU reveals that (i) suppressing the sensitive prefix alone is sufficient to sever the causal generation link, and (ii) flattening only the top-$k$ logits is adequate to maximize uncertainty in the critical subspace. These findings allow PALU to avoid redundant optimization across the full vocabulary and parameter space while minimizing collateral damage to general model performance. Extensive experiments validate that PALU achieves superior forgetting efficacy and utility preservation compared to state-of-the-art baselines.",
      "authors": [
        "Naixin Zhai",
        "Pengyang Shao",
        "Binbin Zheng",
        "Yonghui Yang",
        "Fei Shen",
        "Long Bai",
        "Xun Yang"
      ],
      "published_at": "2026-01-06T17:10:48",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03190v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 5,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4937289357185364,
      "sim_memory": 0.5283973217010498,
      "sim_negative": 0.2787211835384369,
      "broad_relevance": 0.4937289357185364,
      "momentum": 0.06108604879161034,
      "teachability": 0.85,
      "novelty_score": 0.5714424252510071,
      "evidence_score": 0.6091759469228055,
      "direct_memory_relevance": 0.5283973217010498,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8599999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5929286462999549,
      "memory_score": 0.05602031014105155,
      "quality_score": 0.6666703787691222,
      "bridge_score": 0.05602031014105155,
      "max_axis_score": 0.5929286462999549,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "LLM unlearning is an active topic because labs increasingly need practical ways to remove sensitive knowledge without wrecking model utility. This paper offers a sharper recipe: only disrupt the sensitive prefix and only flatten the top-k logits, which is an intuitively teachable refinement over blunt global unlearning.",
      "why_not_higher": "It is not a memory/storage paper in the podcast's first-class sense, and its main contribution is a more targeted unlearning objective rather than a broadly system-changing training or serving method. The abstract claims strong results, but does not clearly establish realism across frontier-scale models or deployment settings.",
      "downgrade_reasons": [
        "memory/storage link absent",
        "primarily an unlearning-method paper",
        "unclear validation on frontier-scale models",
        "may be incremental relative to recent selective-editing/unlearning work"
      ],
      "what_would_raise_priority": "Evidence on larger production-relevant models plus clear comparisons on compute cost, stability, and transfer to real deletion workflows would raise priority.",
      "one_sentence_episode_hook": "What if effective LLM unlearning doesn't require scrambling everything\u2014just the sensitive prefix and a tiny top-k slice of the output distribution?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03233",
      "title": "LTX-2: Efficient Joint Audio-Visual Foundation Model",
      "abstract": "Recent text-to-video diffusion models can generate compelling video sequences, yet they remain silent -- missing the semantic, emotional, and atmospheric cues that audio provides. We introduce LTX-2, an open-source foundational model capable of generating high-quality, temporally synchronized audiovisual content in a unified manner. LTX-2 consists of an asymmetric dual-stream transformer with a 14B-parameter video stream and a 5B-parameter audio stream, coupled through bidirectional audio-video cross-attention layers with temporal positional embeddings and cross-modality AdaLN for shared timestep conditioning. This architecture enables efficient training and inference of a unified audiovisual model while allocating more capacity for video generation than audio generation. We employ a multilingual text encoder for broader prompt understanding and introduce a modality-aware classifier-free guidance (modality-CFG) mechanism for improved audiovisual alignment and controllability. Beyond generating speech, LTX-2 produces rich, coherent audio tracks that follow the characters, environment, style, and emotion of each scene -- complete with natural background and foley elements. In our evaluations, the model achieves state-of-the-art audiovisual quality and prompt adherence among open-source systems, while delivering results comparable to proprietary models at a fraction of their computational cost and inference time. All model weights and code are publicly released.",
      "authors": [
        "Yoav HaCohen",
        "Benny Brazowski",
        "Nisan Chiprut",
        "Yaki Bitterman",
        "Andrew Kvochko",
        "Avishai Berkowitz",
        "Daniel Shalem",
        "Daphna Lifschitz",
        "Dudu Moshe",
        "Eitan Porat"
      ],
      "published_at": "2026-01-06T18:24:41",
      "categories": [
        "cs.CV"
      ],
      "url": "http://arxiv.org/abs/2601.03233v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 31,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "vision",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.39005953073501587,
      "sim_memory": 0.3181798458099365,
      "sim_negative": 0.3072802722454071,
      "broad_relevance": 0.39005953073501587,
      "momentum": 0.2604762596823727,
      "teachability": 0.6,
      "novelty_score": 0.5765587687492371,
      "evidence_score": 0.6799999999999999,
      "direct_memory_relevance": 0.3181798458099365,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.43000000000000005,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.6005969264693648,
      "memory_score": 0.15445395374298096,
      "quality_score": 0.595,
      "bridge_score": 0.15445395374298096,
      "max_axis_score": 0.6005969264693648,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.0315,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Inference",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "Open-source, jointly generated audio-video foundation models are timely because the field is moving from silent video generation to full multimodal media generation. A broad AI audience will care about whether unified audiovisual generation can approach proprietary systems while being practical to run and inspect.",
      "why_not_higher": "The main contribution is a strong multimodal generative model rather than a broadly reusable systems idea, and the memory/storage angle is weak. It also risks being covered as a product-capability milestone more than a paper that changes how most AI teams build systems.",
      "downgrade_reasons": [
        "memory/storage connection is indirect",
        "primarily a model-and-benchmark advance",
        "unclear breadth of transfer beyond audiovisual generation",
        "evidence may be strong within domain but not clearly transformative for general AI systems"
      ],
      "what_would_raise_priority": "Clear ablations and deployment evidence showing that the asymmetric dual-stream design materially changes training or serving efficiency for large multimodal models beyond this specific audiovisual task would raise priority.",
      "one_sentence_episode_hook": "What does it take to make open-source video models stop being silent\u2014and can a single asymmetric transformer generate both the scene and its soundtrack without blowing up cost?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03164",
      "title": "WebAnchor: Anchoring Agent Planning to Stabilize Long-Horizon Web Reasoning",
      "abstract": "Large Language Model(LLM)-based agents have shown strong capabilities in web information seeking, with reinforcement learning (RL) becoming a key optimization paradigm. However, planning remains a bottleneck, as existing methods struggle with long-horizon strategies. Our analysis reveals a critical phenomenon, plan anchor, where the first reasoning step disproportionately impacts downstream behavior in long-horizon web reasoning tasks. Current RL algorithms, fail to account for this by uniformly distributing rewards across the trajectory. To address this, we propose Anchor-GRPO, a two-stage RL framework that decouples planning and execution. In Stage 1, the agent optimizes its first-step planning using fine-grained rubrics derived from self-play experiences and human calibration. In Stage 2, execution is aligned with the initial plan through sparse rewards, ensuring stable and efficient tool usage. We evaluate Anchor-GRPO on four benchmarks: BrowseComp, BrowseComp-Zh, GAIA, and XBench-DeepSearch. Across models from 3B to 30B, Anchor-GRPO outperforms baseline GRPO and First-step GRPO, improving task success and tool efficiency. Notably, WebAnchor-30B achieves 46.0% pass@1 on BrowseComp and 76.4% on GAIA. Anchor-GRPO also demonstrates strong scalability, getting higher accuracy as model size and context length increase.",
      "authors": [
        "Xinmiao Yu",
        "Liwen Zhang",
        "Xiaocheng Feng",
        "Yong Jiang",
        "Bing Qin",
        "Pengjun Xie",
        "Jingren Zhou"
      ],
      "published_at": "2026-01-06T16:36:40",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03164v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 2,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5007184147834778,
      "sim_memory": 0.32674306631088257,
      "sim_negative": 0.26186734437942505,
      "broad_relevance": 0.5007184147834778,
      "momentum": 0.031333575283049045,
      "teachability": 0.85,
      "novelty_score": 0.5116108357906342,
      "evidence_score": 0.559861228866811,
      "direct_memory_relevance": 0.32674306631088257,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.58970304919027,
      "memory_score": 0.0272062673572821,
      "quality_score": 0.6389444915467244,
      "bridge_score": 0.0272062673572821,
      "max_axis_score": 0.58970304919027,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Inference",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "Web agents and long-horizon reasoning are active topics, and this paper offers a concrete training idea: explicitly optimize the first planning step because it anchors the whole trajectory. That is timely for anyone trying to make browser or tool-using agents more reliable.",
      "why_not_higher": "This is not a memory/storage paper in any meaningful systems sense, and its core contribution is still fairly specific to web reasoning agents trained with RL. The claimed insight is interesting, but it does not yet clearly change mainstream LLM building or serving outside this niche.",
      "downgrade_reasons": [
        "memory/storage link is absent",
        "agent-RL niche rather than broad systems impact",
        "benchmark-driven with limited proof of generality beyond web tasks",
        "does not clearly alter serving stack or model architecture"
      ],
      "what_would_raise_priority": "Show the anchor effect and the two-stage training scheme transferring robustly to broader tool-use and agent settings beyond web benchmarks, ideally with stronger ablations and external replications.",
      "one_sentence_episode_hook": "What if the fate of a long-running web agent is mostly decided by its very first thought\u2014and RL has been rewarding the wrong part of the trajectory all along?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02972",
      "title": "Correct, Concise and Complete: Multi-stage Training For Adaptive Reasoning",
      "abstract": "The reasoning capabilities of large language models (LLMs) have improved substantially through increased test-time computation, typically in the form of intermediate tokens known as chain-of-thought (CoT). However, CoT often becomes unnecessarily long, increasing computation cost without actual accuracy gains or sometimes even degrading performance, a phenomenon known as ``overthinking''. We propose a multi-stage efficient reasoning method that combines supervised fine-tuning -- via rejection sampling or reasoning trace reformatting -- with reinforcement learning using an adaptive length penalty. We introduce a lightweight reward function that penalizes tokens generated after the first correct answer but encouraging self-verification only when beneficial. We conduct a holistic evaluation across seven diverse reasoning tasks, analyzing the accuracy-response length trade-off. Our approach reduces response length by an average of 28\\% for 8B models and 40\\% for 32B models, while incurring only minor performance drops of 1.6 and 2.5 points, respectively. Despite its conceptual simplicity, it achieves a superior trade-off compared to more complex state-of-the-art efficient reasoning methods, scoring 76.6, in terms of the area under the Overthinking-Adjusted Accuracy curve ($\\text{AUC}_{\\text{OAA}}$) -- 5 points above the base model and 2.5 points above the second-best approach.",
      "authors": [
        "Nathana\u00ebl Carraz Rakotonirina",
        "Ren Pang",
        "Neha Anna John",
        "Michael Bohlke-Schneider",
        "Momchil Hardalov"
      ],
      "published_at": "2026-01-06T12:31:51",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.02972v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4297025799751282,
      "sim_memory": 0.41212713718414307,
      "sim_negative": 0.19631147384643555,
      "broad_relevance": 0.4297025799751282,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5553911328315735,
      "evidence_score": 0.5493147180559945,
      "direct_memory_relevance": 0.41212713718414307,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5961148418185735,
      "memory_score": 0.10795590732196225,
      "quality_score": 0.5727258872223978,
      "bridge_score": 0.10795590732196225,
      "max_axis_score": 0.5961148418185735,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Efficient reasoning and 'overthinking' are active concerns for LLM builders because test-time compute costs are rising and users increasingly care about latency-cost-quality tradeoffs. This paper offers a simple training recipe to shorten chain-of-thought while mostly preserving accuracy across several reasoning tasks and model sizes.",
      "why_not_higher": "The contribution is more about response-length control than a direct systems or memory mechanism, so it is only indirectly relevant to the podcast's memory/storage lens. Its gains are measured mostly as token reduction with modest accuracy loss, rather than a clear change in deployed serving architecture or hardware behavior.",
      "downgrade_reasons": [
        "memory connection is indirect rather than core",
        "primarily an efficient reasoning/behavior shaping paper",
        "unclear impact on real serving stacks beyond fewer output tokens",
        "incremental relative to broader test-time compute optimization trend"
      ],
      "what_would_raise_priority": "Evidence that the method materially improves real-world serving cost, latency, or KV-cache footprint in production-style inference pipelines would raise priority.",
      "one_sentence_episode_hook": "Can you train an LLM to stop thinking right after it knows the answer\u2014without making it dumber?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03220",
      "title": "From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence",
      "abstract": "Can we learn more from data than existed in the generating process itself? Can new and useful information be constructed from merely applying deterministic transformations to existing data? Can the learnable content in data be evaluated without considering a downstream task? On these questions, Shannon information and Kolmogorov complexity come up nearly empty-handed, in part because they assume observers with unlimited computational capacity and do not target the useful information content. In this work, we identify and exemplify three seeming paradoxes in information theory: (1) information cannot be increased by deterministic transformations; (2) information is independent of the order of data; (3) likelihood modeling is merely distribution matching. To shed light on the tension between these results and modern practice, and to quantify the value of data, we introduce epiplexity, a formalization of information capturing what computationally bounded observers can learn from data. Epiplexity captures the structural content in data while excluding time-bounded entropy, the random unpredictable content exemplified by pseudorandom number generators and chaotic dynamical systems. With these concepts, we demonstrate how information can be created with computation, how it depends on the ordering of the data, and how likelihood modeling can produce more complex programs than present in the data generating process itself. We also present practical procedures to estimate epiplexity which we show capture differences across data sources, track with downstream performance, and highlight dataset interventions that improve out-of-distribution generalization. In contrast to principles of model selection, epiplexity provides a theoretical foundation for data selection, guiding how to select, generate, or transform data for learning systems.",
      "authors": [
        "Marc Finzi",
        "Shikai Qiu",
        "Yiding Jiang",
        "Pavel Izmailov",
        "J. Zico Kolter",
        "Andrew Gordon Wilson"
      ],
      "published_at": "2026-01-06T18:04:03",
      "categories": [
        "cs.LG",
        "stat.ML"
      ],
      "url": "http://arxiv.org/abs/2601.03220v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 9,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.32375603914260864,
      "sim_memory": 0.251987099647522,
      "sim_negative": 0.14559954404830933,
      "broad_relevance": 0.32375603914260864,
      "momentum": 0.1922221879766786,
      "teachability": 0.85,
      "novelty_score": 0.6293531656265259,
      "evidence_score": 0.5702585092994046,
      "direct_memory_relevance": 0.251987099647522,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.82,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5745130005770078,
      "memory_score": 0.046227151010185125,
      "quality_score": 0.6671034037197617,
      "bridge_score": 0.046227151010185125,
      "max_axis_score": 0.5745130005770078,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [
        "compound_window:90d"
      ],
      "time_window": "90d",
      "compound_window_boost": 0.0315,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Theory"
      ],
      "status": "Monitor",
      "why_now": "It challenges core intuitions from Shannon-style information theory at a moment when the field is actively rethinking what data quality and learnability really mean for foundation models. The proposed notion of epiplexity could become a useful framing for data selection and dataset transformation if it holds up.",
      "why_not_higher": "The paper is primarily conceptual and theoretical, with limited direct guidance on how practitioners should change model training or serving today. It also has no concrete memory/storage systems mechanism despite superficial overlap with terms like information and complexity.",
      "downgrade_reasons": [
        "theory-heavy with unclear near-term practice change",
        "memory/storage relevance is absent rather than direct",
        "evidence appears indirect rather than on realistic large-scale systems",
        "editorial fit is stronger for foundations than for transformer systems"
      ],
      "what_would_raise_priority": "Clear demonstrations that epiplexity-based data selection materially improves modern LLM or multimodal training pipelines at realistic scale would raise it substantially.",
      "one_sentence_episode_hook": "What if today's information theory is missing the very thing AI systems actually learn from data once computation is bounded?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03444",
      "title": "Grading Scale Impact on LLM-as-a-Judge: Human-LLM Alignment Is Highest on 0-5 Grading Scale",
      "abstract": "Large language models (LLMs) are increasingly used as automated evaluators, yet prior works demonstrate that these LLM judges often lack consistency in scoring when the prompt is altered. However, the effect of the grading scale itself remains underexplored. We study the LLM-as-a-judge problem by comparing two kinds of raters: humans and LLMs. We collect ratings from both groups on three scales and across six benchmarks that include objective, open-ended subjective, and mixed tasks. Using intraclass correlation coefficients (ICC) to measure absolute agreement, we find that LLM judgments are not perfectly consistent across scales on subjective benchmarks, and that the choice of scale substantially shifts human-LLM agreement, even when within-group panel reliability is high. Aggregated over tasks, the grading scale of 0-5 yields the strongest human-LLM alignment. We further demonstrate that pooled reliability can mask benchmark heterogeneity and reveal systematic subgroup differences in alignment across gender groups, strengthening the importance of scale design and sub-level diagnostics as essential components of LLM-as-a-judge protocols.",
      "authors": [
        "Weiyue Li",
        "Minda Zhao",
        "Weixuan Dong",
        "Jiahui Cai",
        "Yuze Wei",
        "Michael Pocress",
        "Yi Li",
        "Wanyan Yuan",
        "Xiaoyue Wang",
        "Ruoyu Hou"
      ],
      "published_at": "2026-01-06T22:12:06",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.HC"
      ],
      "url": "http://arxiv.org/abs/2601.03444v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 4,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.504539966583252,
      "sim_memory": 0.2526509463787079,
      "sim_negative": 0.1638200879096985,
      "broad_relevance": 0.504539966583252,
      "momentum": 0.05256382402428469,
      "teachability": 0.75,
      "novelty_score": 0.533791184425354,
      "evidence_score": 0.6209437912434099,
      "direct_memory_relevance": 0.2526509463787079,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5860850011301472,
      "memory_score": 0.03810853886282156,
      "quality_score": 0.5893775164973639,
      "bridge_score": 0.03810853886282156,
      "max_axis_score": 0.5860850011301472,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "LLM-as-a-judge is rapidly becoming part of model evaluation pipelines, and this paper identifies a simple prompt-design variable\u2014the grading scale\u2014that materially changes human-LLM agreement. That is timely for anyone using automated judges in benchmarks, product evals, or safety reviews.",
      "why_not_higher": "The contribution is mainly an evaluation-protocol finding rather than a change to how models are trained or served. It also has no real memory/storage or systems angle, and the headline result is a benchmark-design recommendation rather than a broader technical advance.",
      "downgrade_reasons": [
        "No direct memory/storage relevance",
        "Primarily an evaluation study, not a systems or modeling advance",
        "Effect may depend on benchmark and subjective-task mix",
        "More protocol guidance than enduring method"
      ],
      "what_would_raise_priority": "Show that the 0-5 scale recommendation reliably improves downstream ranking stability and decision quality across major judge models and real production evaluation workflows.",
      "one_sentence_episode_hook": "If changing nothing but the score scale can shift how well GPT-style judges agree with humans, how much of today's LLM evaluation stack is built on a formatting accident?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03315",
      "title": "Why LLMs Aren't Scientists Yet: Lessons from Four Autonomous Research Attempts",
      "abstract": "We report a case study of four end-to-end attempts to autonomously generate ML research papers using a pipeline of six LLM agents mapped to stages of the scientific workflow. Of these four, three attempts failed during implementation or evaluation. One completed the pipeline and was accepted to Agents4Science 2025, an experimental inaugural venue that required AI systems as first authors, passing both human and multi-AI review. From these attempts, we document six recurring failure modes: bias toward training data defaults, implementation drift under execution pressure, memory and context degradation across long-horizon tasks, overexcitement that declares success despite obvious failures, insufficient domain intelligence, and weak scientific taste in experimental design. We conclude by discussing four design principles for more robust AI-scientist systems, implications for autonomous scientific discovery, and we release all prompts, artifacts, and outputs at https://github.com/Lossfunk/ai-scientist-artefacts-v1",
      "authors": [
        "Dhruv Trehan",
        "Paras Chopra"
      ],
      "published_at": "2026-01-06T13:20:54",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03315v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 2,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.4818698465824127,
      "sim_memory": 0.3229925334453583,
      "sim_negative": 0.24822278320789337,
      "broad_relevance": 0.4818698465824127,
      "momentum": 0.031333575283049045,
      "teachability": 0.75,
      "novelty_score": 0.4719316363334656,
      "evidence_score": 0.44986122886681096,
      "direct_memory_relevance": 0.3229925334453583,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5880965988113751,
      "memory_score": 0.09008110749762474,
      "quality_score": 0.5689444915467243,
      "bridge_score": 0.09008110749762474,
      "max_axis_score": 0.5880965988113751,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Systems"
      ],
      "status": "Monitor",
      "why_now": "There is strong current interest in autonomous AI research agents, and a negative case study on why they fail is more useful than another hype demo. The paper offers concrete failure modes and design lessons that a broad AI audience can discuss immediately.",
      "why_not_higher": "The evidence base is small: four attempts, one acceptance at a niche experimental venue, and limited proof that the lessons generalize across labs, models, or domains. Its memory discussion is about agent context degradation, not memory/storage systems in the podcast's first-class sense.",
      "downgrade_reasons": [
        "small-n case study",
        "limited external validation",
        "memory is agent-context related, not systems-memory",
        "does not directly change training or serving stacks",
        "venue signal is weak"
      ],
      "what_would_raise_priority": "A broader comparative study across stronger models, longer-horizon tasks, and multiple autonomous-research frameworks with quantified interventions would raise it.",
      "one_sentence_episode_hook": "What happens when you actually let LLM agents try to do science end to end\u2014and three out of four crash before the paper is real?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03464",
      "title": "Prompting Underestimates LLM Capability for Time Series Classification",
      "abstract": "Prompt-based evaluations suggest that large language models (LLMs) perform poorly on time series classification, raising doubts about whether they encode meaningful temporal structure. We show that this conclusion reflects limitations of prompt-based generation rather than the model's representational capacity by directly comparing prompt outputs with linear probes over the same internal representations. While zero-shot prompting performs near chance, linear probes improve average F1 from 0.15-0.26 to 0.61-0.67, often matching or exceeding specialized time series models. Layer-wise analyses further show that class-discriminative time series information emerges in early transformer layers and is amplified by visual and multimodal inputs. Together, these results demonstrate a systematic mismatch between what LLMs internally represent and what prompt-based evaluation reveals, leading current evaluations to underestimate their time series understanding.",
      "authors": [
        "Dan Schumacher",
        "Erfan Nourbakhsh",
        "Rocky Slavin",
        "Anthony Rios"
      ],
      "published_at": "2026-01-06T23:31:27",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03464v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4599223732948303,
      "sim_memory": 0.37447041273117065,
      "sim_negative": 0.36967337131500244,
      "broad_relevance": 0.4599223732948303,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5471685528755188,
      "evidence_score": 0.5093147180559946,
      "direct_memory_relevance": 0.37447041273117065,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.62,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.583947392821076,
      "memory_score": 0.06365888998607055,
      "quality_score": 0.5527258872223978,
      "bridge_score": 0.06365888998607055,
      "max_axis_score": 0.583947392821076,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "It pushes on a live question for foundation-model evaluation: whether prompting is a faithful test of what LLMs actually represent. The probe-vs-prompt mismatch is broadly relevant as people increasingly use generative prompting as a universal evaluation interface.",
      "why_not_higher": "The core result is on time series classification, which is still a relatively narrow downstream task rather than a broad shift in how most teams build or serve LLM systems. It also does not present a direct memory/storage mechanism or systems contribution.",
      "downgrade_reasons": [
        "narrow task focus",
        "no direct memory/storage contribution",
        "evaluation insight more than deployment impact",
        "limited evidence of realistic production workloads"
      ],
      "what_would_raise_priority": "Show the same prompt-vs-representation gap across multiple non-text modalities or mainstream LLM evaluation settings with clear implications for model selection and deployment.",
      "one_sentence_episode_hook": "What if LLMs understand time series much better than prompting suggests\u2014and our favorite evaluation method is the thing that's broken?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03425",
      "title": "The Illusion of Specialization: Unveiling the Domain-Invariant \"Standing Committee\" in Mixture-of-Experts Models",
      "abstract": "Mixture of Experts models are widely assumed to achieve domain specialization through sparse routing. In this work, we question this assumption by introducing COMMITTEEAUDIT, a post hoc framework that analyzes routing behavior at the level of expert groups rather than individual experts. Across three representative models and the MMLU benchmark, we uncover a domain-invariant Standing Committee. This is a compact coalition of routed experts that consistently captures the majority of routing mass across domains, layers, and routing budgets, even when architectures already include shared experts. Qualitative analysis further shows that Standing Committees anchor reasoning structure and syntax, while peripheral experts handle domain-specific knowledge. These findings reveal a strong structural bias toward centralized computation, suggesting that specialization in Mixture of Experts models is far less pervasive than commonly believed. This inherent bias also indicates that current training objectives, such as load-balancing losses that enforce uniform expert utilization, may be working against the model's natural optimization path, thereby limiting training efficiency and performance.",
      "authors": [
        "Yan Wang",
        "Yitao Xu",
        "Nanhan Shen",
        "Jinyan Su",
        "Jimin Huang",
        "Zining Zhu"
      ],
      "published_at": "2026-01-06T21:29:45",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03425v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5855069756507874,
      "sim_memory": 0.5831314921379089,
      "sim_negative": 0.20775866508483887,
      "broad_relevance": 0.5855069756507874,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.4209146499633789,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.5831314921379089,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5687892901897431,
      "memory_score": 0.1759394476413727,
      "quality_score": 0.575,
      "bridge_score": 0.1759394476413727,
      "max_axis_score": 0.5687892901897431,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Inference",
        "Theory"
      ],
      "status": "Monitor",
      "why_now": "MoE models are central to current frontier model scaling, and this paper challenges the widely repeated story that sparse routing naturally yields strong domain specialization. If the reported 'standing committee' effect is real, it has immediate implications for how people interpret expert utilization and design routing or load-balancing objectives.",
      "why_not_higher": "The work is more interpretive and diagnostic than actionable, and the abstract does not show a concrete change to serving efficiency, memory behavior, or training recipes on production-scale workloads. Its memory relevance is indirect: centralized expert usage may matter for systems later, but the paper itself is not about KV cache, bandwidth, offload, or storage mechanisms.",
      "downgrade_reasons": [
        "memory/storage link is indirect rather than core",
        "primarily analysis of routing behavior, not a demonstrated systems improvement",
        "evidence appears benchmark-centric (MMLU) from the abstract",
        "unclear validation on frontier-scale production MoE deployments"
      ],
      "what_would_raise_priority": "A stronger priority case would come from showing that exploiting the standing committee changes MoE training objectives or serving layouts with measurable quality, efficiency, or memory/bandwidth gains at scale.",
      "one_sentence_episode_hook": "What if MoE models are not really a crowd of specialists, but a small standing committee doing most of the work while the rest mostly decorate the illusion?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "memory": [
    {
      "arxiv_id": "2601.03197",
      "title": "Software-Defined Agentic Serving",
      "abstract": "As multi-agent LLM pipelines grow in complexity, existing serving paradigms fail to adapt to the dynamic serving conditions. We argue that agentic serving systems should be programmable and system-aware, unlike existing serving which statically encode the parameters. In this work, we propose a new SDN-inspired agentic serving framework that helps control the key attributes of communication based on runtime state. This architecture enables serving-efficient, responsive agent systems and paves the way for high-level intent-driven agentic serving.",
      "authors": [
        "Saurabh Agarwal",
        "Marco Laju",
        "Jayanth Srinivasa",
        "Myungjin Lee",
        "Aditya Akella"
      ],
      "published_at": "2026-01-06T17:22:45",
      "categories": [
        "cs.DC",
        "cs.MA"
      ],
      "url": "http://arxiv.org/abs/2601.03197v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.4616243243217468,
      "sim_memory": 0.3901659846305847,
      "sim_negative": 0.16224351525306702,
      "broad_relevance": 0.4616243243217468,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.40322721004486084,
      "evidence_score": 0.28,
      "direct_memory_relevance": 0.3901659846305847,
      "systems_leverage": 0.7,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.83,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.38897137880325316,
      "memory_score": 0.2560497953891754,
      "quality_score": 0.595,
      "bridge_score": 0.2560497953891754,
      "max_axis_score": 0.38897137880325316,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Systems",
        "Inference",
        "Memory/Storage Adjacent",
        "Memory-relevant"
      ],
      "status": "Monitor",
      "why_now": "Agentic LLM pipelines are getting more complex, and serving them under changing runtime conditions is a real emerging systems problem. A programmable, SDN-inspired control plane for agent communication is timely as teams move from single-model inference to multi-step agent orchestration.",
      "why_not_higher": "From the abstract, the memory/storage angle is indirect rather than central, and the evidence is unclear on realistic models, workloads, and hardware. It also sounds more like an architecture proposal than a demonstrated shift in how most practitioners will serve AI in the next 30-90 days.",
      "downgrade_reasons": [
        "memory connection is adjacent not core",
        "abstract emphasizes framework vision over concrete results",
        "unclear realism of workloads and deployment hardware",
        "multi-agent serving is important but still niche for broad audience"
      ],
      "what_would_raise_priority": "Strong end-to-end results on realistic multi-agent LLM workloads showing clear gains in latency, throughput, cost, or bandwidth/data-movement efficiency versus current serving stacks would raise it.",
      "one_sentence_episode_hook": "What if serving AI agents looked less like static inference and more like software-defined networking with a control plane that adapts communication on the fly?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.11595",
      "title": "Enhancing Model Context Protocol (MCP) with Context-Aware Server Collaboration",
      "abstract": "The Model Context Protocol (MCP) (MCP Community, 2025) has emerged as a widely used framework for enabling LLM-based agents to communicate with external tools and services. The original MCP implementation (Anthropic, 2024) relies on a Large Language Model (LLM) to decompose tasks and issue instructions to servers. In particular, the agents, models, and servers are stateless and do not have access to a global context. However, in tasks involving LLM-driven coordination, it is natural that a Shared Context Store (SCS) could improve the efficiency and coherence of multi-agent workflows by reducing redundancy and enabling knowledge transfer between servers. Thus, in this work, we design and assess the performance of a Context-Aware MCP (CA-MCP) that offloads execution logic to specialized MCP servers that read from and write to a shared context memory, allowing them to coordinate more autonomously in real time. In this design, context management serves as the central mechanism that maintains continuity across task executions by tracking intermediate states and shared variables, thereby enabling persistent collaboration among agents without repeated prompting. We present experiments showing that the CA-MCP can outperform the traditional MCP by reducing the number of LLM calls required for complex tasks and decreasing the frequency of response failures when task conditions are not satisfied. In particular, we conducted experiments on the TravelPlanner (Yang et al., 2024) and REALM-Bench (Geng & Chang, 2025) benchmark datasets and observed statistically significant results indicating the potential advantages of incorporating a shared context store via CA-MCP in LLM-driven multi-agent systems.",
      "authors": [
        "Meenakshi Amulya Jayanti",
        "X. Y. Han"
      ],
      "published_at": "2026-01-06T21:34:08",
      "categories": [
        "cs.DC",
        "cs.LG",
        "cs.SE"
      ],
      "url": "http://arxiv.org/abs/2601.11595v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.444566935300827,
      "sim_memory": 0.3773147463798523,
      "sim_negative": 0.18660494685173035,
      "broad_relevance": 0.444566935300827,
      "momentum": 0.017490950964499406,
      "teachability": 0.75,
      "novelty_score": 0.41364002227783203,
      "evidence_score": 0.4293147180559946,
      "direct_memory_relevance": 0.3773147463798523,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.7799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.38931148183322206,
      "memory_score": 0.18951219008067502,
      "quality_score": 0.6027258872223977,
      "bridge_score": 0.18951219008067502,
      "max_axis_score": 0.38931148183322206,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Adjacent",
        "Systems",
        "Inference",
        "Memory-relevant"
      ],
      "status": "Monitor",
      "why_now": "MCP and tool-using agent stacks are hot, and this paper argues that adding a shared context store can reduce redundant LLM calls and improve coordination failures in multi-agent workflows. That is timely for builders experimenting with agent orchestration beyond pure prompt-driven routing.",
      "why_not_higher": "The memory angle is mostly architectural shared state, not first-class AI memory/storage systems like KV cache, paging, offload, or bandwidth reduction. Evidence appears benchmark-level and workflow-specific rather than a decisive change in how mainstream AI systems are built or served.",
      "downgrade_reasons": [
        "memory connection is indirect rather than systems-core",
        "focuses on MCP agent orchestration rather than core model/training/serving stack",
        "benchmark evidence on task suites, not production deployments or hardware-level behavior",
        "incremental extension of an emerging protocol"
      ],
      "what_would_raise_priority": "A stronger real-world study showing broad gains across diverse agent stacks, models, and production workloads\u2014especially with clear cost, latency, and reliability improvements\u2014would raise it.",
      "one_sentence_episode_hook": "Can a shared context store make MCP agents smarter and cheaper by replacing repeated prompting with real coordination memory?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03236",
      "title": "MAGMA: A Multi-Graph based Agentic Memory Architecture for AI Agents",
      "abstract": "Memory-Augmented Generation (MAG) extends Large Language Models with external memory to support long-context reasoning, but existing approaches largely rely on semantic similarity over monolithic memory stores, entangling temporal, causal, and entity information. This design limits interpretability and alignment between query intent and retrieved evidence, leading to suboptimal reasoning accuracy. In this paper, we propose MAGMA, a multi-graph agentic memory architecture that represents each memory item across orthogonal semantic, temporal, causal, and entity graphs. MAGMA formulates retrieval as policy-guided traversal over these relational views, enabling query-adaptive selection and structured context construction. By decoupling memory representation from retrieval logic, MAGMA provides transparent reasoning paths and fine-grained control over retrieval. Experiments on LoCoMo and LongMemEval demonstrate that MAGMA consistently outperforms state-of-the-art agentic memory systems in long-horizon reasoning tasks.",
      "authors": [
        "Dongming Jiang",
        "Yi Li",
        "Guanpeng Li",
        "Bingzhe Li"
      ],
      "published_at": "2026-01-06T18:29:43",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03236v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 5,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.44615691900253296,
      "sim_memory": 0.34826526045799255,
      "sim_negative": 0.21989858150482178,
      "broad_relevance": 0.44615691900253296,
      "momentum": 0.06108604879161034,
      "teachability": 0.85,
      "novelty_score": 0.5349092781543732,
      "evidence_score": 0.5391759469228055,
      "direct_memory_relevance": 0.34826526045799255,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.62,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5431770692206588,
      "memory_score": 0.12898069176813443,
      "quality_score": 0.6166703787691222,
      "bridge_score": 0.12898069176813443,
      "max_axis_score": 0.5431770692206588,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Memory/Storage Adjacent",
        "Inference",
        "Memory-relevant"
      ],
      "status": "Monitor",
      "why_now": "Agent memory is a live topic as more teams try to make LLM agents handle long-horizon tasks, and this paper offers a cleaner retrieval formulation using multiple relational graphs instead of a single semantic store. The interpretability angle and reported gains on long-memory benchmarks make it timely for the current agentic-memory discussion.",
      "why_not_higher": "This is memory in the agent-retrieval sense, not memory/storage in the systems sense the show prioritizes, and the abstract does not show concrete implications for KV cache, bandwidth, offload, or serving. Evidence appears benchmark-centered rather than demonstrated on production-scale workloads, models, or hardware constraints.",
      "downgrade_reasons": [
        "agent memory rather than systems memory",
        "no clear cache/offload/bandwidth mechanism",
        "benchmark-focused evidence",
        "unclear deployment impact on modern serving stacks"
      ],
      "what_would_raise_priority": "A stronger case would include ablations on retrieval cost, latency, scale, and integration with real agent stacks or long-context serving constraints.",
      "one_sentence_episode_hook": "Can AI agents remember better by splitting memory into semantic, temporal, causal, and entity graphs instead of dumping everything into one vector store?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "monitor": [
    {
      "arxiv_id": "2601.03014",
      "title": "SentGraph: Hierarchical Sentence Graph for Multi-hop Retrieval-Augmented Question Answering",
      "abstract": "Traditional Retrieval-Augmented Generation (RAG) effectively supports single-hop question answering with large language models but faces significant limitations in multi-hop question answering tasks, which require combining evidence from multiple documents. Existing chunk-based retrieval often provides irrelevant and logically incoherent context, leading to incomplete evidence chains and incorrect reasoning during answer generation. To address these challenges, we propose SentGraph, a sentence-level graph-based RAG framework that explicitly models fine-grained logical relationships between sentences for multi-hop question answering. Specifically, we construct a hierarchical sentence graph offline by first adapting Rhetorical Structure Theory to distinguish nucleus and satellite sentences, and then organizing them into topic-level subgraphs with cross-document entity bridges. During online retrieval, SentGraph performs graph-guided evidence selection and path expansion to retrieve fine-grained sentence-level evidence. Extensive experiments on four multi-hop question answering benchmarks demonstrate the effectiveness of SentGraph, validating the importance of explicitly modeling sentence-level logical dependencies for multi-hop reasoning.",
      "authors": [
        "Junli Liang",
        "Pengfei Zhou",
        "Wangqiu Zhou",
        "Wenjie Qing",
        "Qi Zhao",
        "Ziwen Wang",
        "Qi Song",
        "Xiangyang Li"
      ],
      "published_at": "2026-01-06T13:39:51",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03014v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5089484453201294,
      "sim_memory": 0.27174055576324463,
      "sim_negative": 0.13754987716674805,
      "broad_relevance": 0.5089484453201294,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5609372854232788,
      "evidence_score": 0.4893147180559946,
      "direct_memory_relevance": 0.27174055576324463,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.74,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5407205243108297,
      "memory_score": 0.045839932895692714,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.045839932895692714,
      "max_axis_score": 0.5407205243108297,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "Multi-hop RAG remains a practical pain point, and this paper offers a clear, interpretable alternative to chunk retrieval by retrieving sentence-level evidence chains. That makes it timely for teams trying to improve factual QA and reasoning over multiple documents without changing the base LLM.",
      "why_not_higher": "The contribution is mainly a task-focused retrieval architecture for multi-hop QA rather than a broadly transformative change to how modern AI systems are built or served. Its connection to the show's memory/storage lens is weak because the 'graph' is about evidence structure, not KV cache, bandwidth, paging, or system memory behavior.",
      "downgrade_reasons": [
        "Weak memory/storage relevance",
        "Primarily a narrow multi-hop QA/RAG method",
        "Unclear impact beyond retrieval benchmarks",
        "More likely an incremental retrieval design than a field-shifting result"
      ],
      "what_would_raise_priority": "Stronger evidence that sentence-graph retrieval generalizes across real production RAG settings, larger models, and diverse enterprise corpora would raise its priority.",
      "one_sentence_episode_hook": "What if multi-hop RAG fails not because LLMs cannot reason, but because chunk retrieval never hands them a coherent evidence path in the first place?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03093",
      "title": "ATLAS: Adaptive Test-Time Latent Steering with External Verifiers for Enhancing LLMs Reasoning",
      "abstract": "Recent work on activation and latent steering has demonstrated that modifying internal representations can effectively guide large language models (LLMs) toward improved reasoning and efficiency without additional training. However, most existing approaches rely on fixed steering policies and static intervention strengths, which limit their robustness across problem instances and often result in over- or under-steering. We propose Adaptive Test-time Latent Steering, called (ATLAS), a task-specific framework that dynamically controls steering decisions at inference time using an external, lightweight latent verifier. Given intermediate hidden states, the verifier predicts the quality of ongoing reasoning and adaptively selects whether and how strongly to apply steering, enabling per-example and per-step adjustment with minimal overhead. To our knowledge, ATLAS is the first method to integrate learned latent verification into test-time steering for enhancing LLMs reasoning. Experiments on multiple mathematical reasoning benchmarks show that ATLAS consistently outperforms both vanilla decoding and fixed steering baselines, achieving higher accuracy while substantially reducing test-time token usage. These results demonstrate that verifier-guided latent adaptation provides an effective and scalable mechanism for controlling reasoning efficiency without sacrificing solution quality. All source code will be publicly available.",
      "authors": [
        "Tuc Nguyen",
        "Thai Le"
      ],
      "published_at": "2026-01-06T15:27:24",
      "categories": [
        "cs.LG",
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03093v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4707915186882019,
      "sim_memory": 0.3888080418109894,
      "sim_negative": 0.22144147753715515,
      "broad_relevance": 0.4707915186882019,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5666391253471375,
      "evidence_score": 0.40931471805599456,
      "direct_memory_relevance": 0.3888080418109894,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7599999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5301287223098302,
      "memory_score": 0.05096017871001615,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.05096017871001615,
      "max_axis_score": 0.5301287223098302,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Test-time control of reasoning is a live topic, and ATLAS adds a potentially useful twist: using a lightweight verifier to decide when and how strongly to steer latent states during inference. If the gains in accuracy and token efficiency hold beyond math benchmarks, this could matter for practical reasoning-time adaptation.",
      "why_not_higher": "The paper appears centered on mathematical reasoning benchmarks rather than broadly realistic production workloads, and the memory/storage angle is not direct. It is also an incremental extension of test-time steering rather than a clear shift in how mainstream systems are built or served.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Evidence seems limited to reasoning benchmarks",
        "Unclear generalization across models and tasks",
        "More method tweak than field-shifting result"
      ],
      "what_would_raise_priority": "Show robust gains across diverse tasks and model families, with clear latency/cost tradeoffs in realistic serving settings.",
      "one_sentence_episode_hook": "Can a tiny external verifier decide, step by step, when an LLM should nudge its own hidden states\u2014and buy better reasoning with fewer tokens?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02989",
      "title": "Mechanistic Interpretability of Large-Scale Counting in LLMs through a System-2 Strategy",
      "abstract": "Large language models (LLMs), despite strong performance on complex mathematical problems, exhibit systematic limitations in counting tasks. This issue arises from architectural limits of transformers, where counting is performed across layers, leading to degraded precision for larger counting problems due to depth constraints. To address this limitation, we propose a simple test-time strategy inspired by System-2 cognitive processes that decomposes large counting tasks into smaller, independent sub-problems that the model can reliably solve. We evaluate this approach using observational and causal mediation analyses to understand the underlying mechanism of this System-2-like strategy. Our mechanistic analysis identifies key components: latent counts are computed and stored in the final item representations of each part, transferred to intermediate steps via dedicated attention heads, and aggregated in the final stage to produce the total count. Experimental results demonstrate that this strategy enables LLMs to surpass architectural limitations and achieve high accuracy on large-scale counting tasks. This work provides mechanistic insight into System-2 counting in LLMs and presents a generalizable approach for improving and understanding their reasoning behavior.",
      "authors": [
        "Hosein Hasani",
        "Mohammadali Banayeeanzade",
        "Ali Nafisi",
        "Sadegh Mohammadian",
        "Fatemeh Askari",
        "Mobin Bagherian",
        "Amirmohammad Izadi",
        "Mahdieh Soleymani Baghshah"
      ],
      "published_at": "2026-01-06T12:58:27",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.02989v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.44576042890548706,
      "sim_memory": 0.39225369691848755,
      "sim_negative": 0.2926170229911804,
      "broad_relevance": 0.44576042890548706,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5747885406017303,
      "evidence_score": 0.40931471805599456,
      "direct_memory_relevance": 0.39225369691848755,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7599999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5438418076632047,
      "memory_score": 0.0689938752422656,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.0689938752422656,
      "max_axis_score": 0.5438418076632047,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Theory",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Mechanistic interpretability plus a simple test-time decomposition strategy is timely because many listeners care about where LLM reasoning fails and whether prompting can bypass architectural limits. Counting is a clean setting for explaining a System-2-like strategy in an educational way.",
      "why_not_higher": "The contribution appears centered on counting, a narrow task, and the abstract does not show clear evidence on realistic production workloads or broad downstream gains. Its memory connection is interpretability-level latent storage inside the model, not a concrete systems paper about cache, bandwidth, or offload.",
      "downgrade_reasons": [
        "narrow counting task",
        "memory/storage link is indirect",
        "unclear realism of workloads and model scale",
        "limited evidence of impact on build/train/serve practice"
      ],
      "what_would_raise_priority": "Show that the same mechanistic decomposition reliably improves broader reasoning tasks or informs practical inference-time methods on widely used frontier or open models.",
      "one_sentence_episode_hook": "Can LLMs beat their own counting limits by turning one hard count into many easy ones\u2014and can mechanistic interpretability show exactly where those partial counts live and how they get added up?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03205",
      "title": "UltraLogic: Enhancing LLM Reasoning through Large-Scale Data Synthesis and Bipolar Float Reward",
      "abstract": "While Large Language Models (LLMs) have demonstrated significant potential in natural language processing , complex general-purpose reasoning requiring multi-step logic, planning, and verification remains a critical bottleneck. Although Reinforcement Learning with Verifiable Rewards (RLVR) has succeeded in specific domains , the field lacks large-scale, high-quality, and difficulty-calibrated data for general reasoning. To address this, we propose UltraLogic, a framework that decouples the logical core of a problem from its natural language expression through a Code-based Solving methodology to automate high-quality data production. The framework comprises hundreds of unique task types and an automated calibration pipeline across ten difficulty levels. Furthermore, to mitigate binary reward sparsity and the Non-negative Reward Trap, we introduce the Bipolar Float Reward (BFR) mechanism, utilizing graded penalties to effectively distinguish perfect responses from those with logical flaws. Our experiments demonstrate that task diversity is the primary driver for reasoning enhancement , and that BFR, combined with a difficulty matching strategy, significantly improves training efficiency, guiding models toward global logical optima.",
      "authors": [
        "Yile Liu",
        "Yixian Liu",
        "Zongwei Li",
        "Yufei Huang",
        "Xinhua Feng",
        "Zhichao Hu",
        "Jinglu Hu",
        "Jianfeng Yan",
        "Fengzong Lian",
        "Yuhong Liu"
      ],
      "published_at": "2026-01-06T17:41:32",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03205v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.439242959022522,
      "sim_memory": 0.29593801498413086,
      "sim_negative": 0.2298593968153,
      "broad_relevance": 0.439242959022522,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5833818316459656,
      "evidence_score": 0.38931471805599455,
      "direct_memory_relevance": 0.29593801498413086,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.85,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5531755603549505,
      "memory_score": 0.040099170661958616,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.040099170661958616,
      "max_axis_score": 0.5531755603549505,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "Reasoning training remains a top audience topic, and this paper packages two timely ideas: large-scale synthetic logic data generation and a less sparse RL-style reward for general reasoning. If the gains hold beyond the authors' setup, it could matter for how labs build post-training reasoning curricula.",
      "why_not_higher": "The paper is not meaningfully about memory or storage systems, and the abstract does not yet show convincing validation on broadly accepted frontier-scale models or production-like workloads. It may be a useful training recipe, but it is still close to a benchmark-and-methods result until external replication or stronger ablations land.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Evidence appears centered on authors' synthetic framework",
        "Unclear realism of workloads and model scale from abstract alone",
        "Crowded reasoning/RLVR landscape"
      ],
      "what_would_raise_priority": "Show robust gains on widely used reasoning benchmarks and strong models, with clear ablations proving BFR and data diversity transfer beyond the UltraLogic task suite.",
      "one_sentence_episode_hook": "Can better synthetic logic data\u2014and a reward signal that punishes almost-right answers\u2014actually teach LLMs to reason better, or is UltraLogic just another reasoning benchmark stack?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.06160",
      "title": "Student Guides Teacher: Weak-to-Strong Inference via Spectral Orthogonal Exploration",
      "abstract": "While Large Language Models (LLMs) demonstrate near-human capabilities, they often suffer from \"Reasoning Collapse\" in complex mathematical proving and long-horizon planning. Models tend to degenerate into low-rank Bias Manifold, where stochastic sampling merely produces lexical variations of erroneous logic rather than semantic exploration. This geometric collapse renders the model \"blind\" to high-value solutions that lie within its Null Space. To address this, we propose Spectral Orthogonal Exploration (SOE), a geometric framework operating on a counter-intuitive \"Student Guides Teacher\" paradigm. Specifically, we utilize a weak auxiliary agent not for imitation, but as an orthogonal probe. By explicitly navigating the Teacher's Null Space, SOE serves as a geometric bridge, effectively ejecting the model from local optima to explore diverse, high-value solution spaces. Experiments on mathematical benchmarks demonstrate that, relative to baseline methods, our approach improves average accuracy by 62.4% and increases average sampling efficiency by 113.7%, indicating a promising path toward overcoming performance plateaus in advanced reasoning tasks.",
      "authors": [
        "Dayu Wang",
        "Jiaye Yang",
        "Weikang Li",
        "Jiahui Liang",
        "Yang Li"
      ],
      "published_at": "2026-01-06T16:39:40",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.06160v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.41302379965782166,
      "sim_memory": 0.37532180547714233,
      "sim_negative": 0.2201598882675171,
      "broad_relevance": 0.41302379965782166,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5628524720668793,
      "evidence_score": 0.3693147180559946,
      "direct_memory_relevance": 0.37532180547714233,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7499999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5122304086086775,
      "memory_score": 0.036914307809862046,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.036914307809862046,
      "max_axis_score": 0.5122304086086775,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Inference",
        "Theory"
      ],
      "status": "Monitor",
      "why_now": "Reasoning-time intervention methods for LLMs are a live topic, and the paper offers a provocative 'student guides teacher' framing for escaping bad solution modes during math-style inference. If the gains hold, it could matter for how people design sampling and search around frontier models without retraining.",
      "why_not_higher": "The abstract leans heavily on geometric language and large relative gains, but it is unclear whether the method is validated on realistic model scales, diverse workloads, and strong baselines beyond math benchmarks. The memory/storage connection is essentially absent, and the practical recipe for broad deployment is not yet convincing from the provided evidence.",
      "downgrade_reasons": [
        "No direct memory or storage mechanism",
        "Evidence appears concentrated on math benchmarks",
        "Unclear realism of workloads and model sizes",
        "Possible benchmark-driven reasoning claim inflation",
        "Transfer to production inference stacks not yet demonstrated"
      ],
      "what_would_raise_priority": "Show robust gains on widely used frontier or open models across multiple reasoning and agentic workloads, with clear ablations against strong inference-time baselines and practical cost tradeoffs.",
      "one_sentence_episode_hook": "What if the best way to make a strong reasoning model think better is to let a weaker model push it out of its favorite mistakes?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02993",
      "title": "Stable-RAG: Mitigating Retrieval-Permutation-Induced Hallucinations in Retrieval-Augmented Generation",
      "abstract": "Retrieval-Augmented Generation (RAG) has become a key paradigm for reducing factual hallucinations in large language models (LLMs), yet little is known about how the order of retrieved documents affects model behavior. We empirically show that under Top-5 retrieval with the gold document included, LLM answers vary substantially across permutations of the retrieved set, even when the gold document is fixed in the first position. This reveals a previously underexplored sensitivity to retrieval permutations. Although robust RAG methods primarily focus on enhancing LLM robustness to low-quality retrieval and mitigating positional bias to distribute attention fairly over long contexts, neither approach directly addresses permutation sensitivity. In this paper, we propose Stable-RAG, which exploits permutation sensitivity estimation to mitigate permutation-induced hallucinations. Stable-RAG runs the generator under multiple retrieval orders, clusters hidden states, and decodes from a cluster-center representation that captures the dominant reasoning pattern. It then uses these reasoning results to align hallucinated outputs toward the correct answer, encouraging the model to produce consistent and accurate predictions across document permutations. Experiments on three QA datasets show that Stable-RAG significantly improves answer accuracy, reasoning consistency and robust generalization across datasets, retrievers, and input lengths compared with baselines.",
      "authors": [
        "Qianchi Zhang",
        "Hainan Zhang",
        "Liang Pang",
        "Hongwei Zheng",
        "Zhiming Zheng"
      ],
      "published_at": "2026-01-06T13:07:38",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.02993v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.5748662948608398,
      "sim_memory": 0.4381977617740631,
      "sim_negative": 0.2910250723361969,
      "broad_relevance": 0.5748662948608398,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.47947949171066284,
      "evidence_score": 0.45,
      "direct_memory_relevance": 0.4381977617740631,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5643818122148514,
      "memory_score": 0.0474593285322189,
      "quality_score": 0.595,
      "bridge_score": 0.0474593285322189,
      "max_axis_score": 0.5643818122148514,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Systems",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "RAG is widely deployed, and this paper highlights a practical failure mode many teams likely have not tested: answer instability under harmless-looking permutations of retrieved documents. A method that improves consistency and accuracy across retrievers and input lengths is timely for anyone shipping retrieval-based LLM systems.",
      "why_not_higher": "The contribution is mainly an inference-time robustness technique for QA-style RAG rather than a broader shift in model training or serving architecture. Its memory/storage relevance is weak, and the approach may add multi-pass generation cost that limits immediate adoption.",
      "downgrade_reasons": [
        "Memory/storage connection is absent",
        "Primarily a RAG evaluation/robustness paper",
        "Potential inference overhead from multiple retrieval orders",
        "Evidence appears task-centered on QA benchmarks rather than broad production workloads"
      ],
      "what_would_raise_priority": "Show strong gains on production-style RAG workloads with clear latency/cost tradeoffs and evidence that the method generalizes beyond QA to mainstream enterprise retrieval settings.",
      "one_sentence_episode_hook": "What if your RAG system hallucinates not because retrieval failed, but because the same documents arrived in a different order?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03448",
      "title": "Enhancing Linguistic Competence of Language Models through Pre-training with Language Learning Tasks",
      "abstract": "Language models (LMs) are pre-trained on raw text datasets to generate text sequences token-by-token. While this approach facilitates the learning of world knowledge and reasoning, it does not explicitly optimize for linguistic competence. To bridge this gap, we propose L2T, a pre-training framework integrating Language Learning Tasks alongside standard next-token prediction. Inspired by human language acquisition, L2T transforms raw text into structured input-output pairs to provide explicit linguistic stimulation. Pre-training LMs on a mixture of raw text and L2T data not only improves overall performance on linguistic competence benchmarks but accelerates its acquisition, while maintaining competitive performance on general reasoning tasks.",
      "authors": [
        "Atsuki Yamaguchi",
        "Maggie Mi",
        "Nikolaos Aletras"
      ],
      "published_at": "2026-01-06T22:28:15",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03448v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4604429602622986,
      "sim_memory": 0.4257631003856659,
      "sim_negative": 0.2211679071187973,
      "broad_relevance": 0.4604429602622986,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.569884181022644,
      "evidence_score": 0.36000000000000004,
      "direct_memory_relevance": 0.4257631003856659,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5236155152320862,
      "memory_score": 0.04872893011569976,
      "quality_score": 0.595,
      "bridge_score": 0.04872893011569976,
      "max_axis_score": 0.5236155152320862,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "It tackles a live question in LLM training: whether next-token prediction alone is enough, or whether explicit linguistic supervision during pre-training can produce better language competence without sacrificing general ability. That makes it potentially interesting to a broad AI audience focused on how to shape pre-training objectives.",
      "why_not_higher": "The paper appears more about benchmarked pre-training recipe changes than a clear shift in how frontier systems are built, and it has no meaningful memory/storage angle. Its practical impact depends on scale, robustness across model families, and whether the added task mixture justifies the extra data construction complexity.",
      "downgrade_reasons": [
        "No direct memory/storage contribution",
        "Primarily a training-objective paper rather than systems change",
        "Unclear evidence on frontier-scale models or production realism",
        "Could land as a benchmark/pretraining-mixture result rather than a durable shift"
      ],
      "what_would_raise_priority": "Stronger evidence that L2T scales to widely used LLM sizes and consistently improves downstream quality-cost tradeoffs across architectures would raise it.",
      "one_sentence_episode_hook": "What if language models learned grammar more like humans do\u2014by mixing next-token prediction with explicit language-learning exercises during pre-training?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03320",
      "title": "Ratio-Variance Regularized Policy Optimization for Efficient LLM Fine-tuning",
      "abstract": "On-policy reinforcement learning (RL), particularly Proximal Policy Optimization (PPO) and Group Relative Policy Optimization (GRPO), has become the dominant paradigm for fine-tuning large language models (LLMs). While policy ratio clipping stabilizes training, this heuristic hard constraint incurs a fundamental cost: it indiscriminately truncates gradients from high-return yet high-divergence actions, suppressing rare but highly informative \"eureka moments\" in complex reasoning. Moreover, once data becomes slightly stale, hard clipping renders it unusable, leading to severe sample inefficiency. In this work, we revisit the trust-region objective in policy optimization and show that explicitly constraining the \\emph{variance (second central moment) of the policy ratio} provides a principled and smooth relaxation of hard clipping. This distributional constraint stabilizes policy updates while preserving gradient signals from valuable trajectories. Building on this insight, we propose $R^2VPO$ (Ratio-Variance Regularized Policy Optimization), a novel primal-dual framework that supports stable on-policy learning and enables principled off-policy data reuse by dynamically reweighting stale samples rather than discarding them. We extensively evaluate $R^2VPO$ on fine-tuning state-of-the-art LLMs, including DeepSeek-Distill-Qwen-1.5B and the openPangu-Embedded series (1B and 7B), across challenging mathematical reasoning benchmarks. Experimental results show that $R^2VPO$ consistently achieves superior asymptotic performance, with average relative gains of up to 17% over strong clipping-based baselines, while requiring approximately 50% fewer rollouts to reach convergence. These findings establish ratio-variance control as a promising direction for improving both stability and data efficiency in RL-based LLM alignment.",
      "authors": [
        "Yu Luo",
        "Shuo Han",
        "Yihan Hu",
        "Dong Li",
        "Jianye Hao"
      ],
      "published_at": "2026-01-06T14:01:42",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03320v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.43891775608062744,
      "sim_memory": 0.46079960465431213,
      "sim_negative": 0.1570451557636261,
      "broad_relevance": 0.43891775608062744,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5350363254547119,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.46079960465431213,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8899999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.521930775642395,
      "memory_score": 0.08123988139629368,
      "quality_score": 0.595,
      "bridge_score": 0.08123988139629368,
      "max_axis_score": 0.521930775642395,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Theory",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "RL fine-tuning for reasoning LLMs is active again, and this paper targets a real pain point in PPO/GRPO-style training: clipping can waste rare high-value trajectories and discard mildly stale data. A smoother trust-region style objective with off-policy reuse is timely if labs are trying to cut rollout cost while improving post-training.",
      "why_not_higher": "This is still mainly an RL objective paper rather than a clear systems or memory/storage paper, and the evidence is benchmark-centric on math reasoning fine-tuning rather than broad production training stacks. The claimed gains are interesting, but it is not yet obvious this will change standard practice across model families and alignment settings.",
      "downgrade_reasons": [
        "Memory/storage link is absent rather than direct",
        "Mostly optimization/alignment methodology, not systems",
        "Evidence centers on a limited set of reasoning benchmarks and model sizes",
        "Unclear adoption path versus entrenched PPO/GRPO pipelines"
      ],
      "what_would_raise_priority": "Independent results on larger frontier-style post-training runs, plus evidence that off-policy reuse materially reduces real compute cost in production RLHF/RLAIF pipelines, would raise it.",
      "one_sentence_episode_hook": "What if PPO\u2019s safety rail is also killing the very rare trajectories that teach LLMs how to reason\u2014and a variance-based trust region can recover them with half the rollouts?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03385",
      "title": "SIGMA: Scalable Spectral Insights for LLM Model Collapse",
      "abstract": "The rapid adoption of synthetic data for training Large Language Models (LLMs) has introduced the technical challenge of \"model collapse\"-a degenerative process where recursive training on model-generated content leads to a contraction of distributional variance and representational quality. While the phenomenology of collapse is increasingly evident, rigorous methods to quantify and predict its onset in high-dimensional spaces remain elusive. In this paper, we introduce SIGMA (Spectral Inequalities for Gram Matrix Analysis), a unified framework that benchmarks model collapse through the spectral lens of the embedding Gram matrix. By deriving and utilizing deterministic and stochastic bounds on the matrix's spectrum, SIGMA provides a mathematically grounded metric to track the contraction of the representation space. Crucially, our stochastic formulation enables scalable estimation of these bounds, making the framework applicable to large-scale foundation models where full eigendecomposition is intractable. We demonstrate that SIGMA effectively captures the transition towards degenerate states, offering both theoretical insights into the mechanics of collapse and a practical, scalable tool for monitoring the health of recursive training pipelines.",
      "authors": [
        "Yi Gu",
        "Lingyou Pang",
        "Xiangkun Ye",
        "Tianyu Wang",
        "Jianyu Lin",
        "Carey E. Priebe",
        "Alexander Aue"
      ],
      "published_at": "2026-01-06T19:47:11",
      "categories": [
        "cs.LG",
        "math.PR"
      ],
      "url": "http://arxiv.org/abs/2601.03385v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5014312267303467,
      "sim_memory": 0.5088726878166199,
      "sim_negative": 0.21531957387924194,
      "broad_relevance": 0.5014312267303467,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.4826890230178833,
      "evidence_score": 0.32,
      "direct_memory_relevance": 0.5088726878166199,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.84,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5228327214717865,
      "memory_score": 0.07566180634498598,
      "quality_score": 0.575,
      "bridge_score": 0.07566180634498598,
      "max_axis_score": 0.5228327214717865,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Theory",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "As labs increasingly recycle synthetic data in post-training and continual training pipelines, a scalable way to detect model collapse is timely and broadly relevant. The spectral framing could give practitioners a concrete health signal for recursive training before visible quality failures appear.",
      "why_not_higher": "The work is primarily a theoretical/diagnostic benchmark rather than a demonstrated change to how frontier models are trained or served today. The memory/storage connection is not direct, and the abstract does not show clear validation on production-scale LLM training pipelines or hardware-constrained systems.",
      "downgrade_reasons": [
        "Memory/storage relevance is absent rather than core",
        "Mostly diagnostic/theoretical framing",
        "Unclear evidence on realistic frontier-scale workloads",
        "Not yet a clear systems or deployment lever"
      ],
      "what_would_raise_priority": "Show that SIGMA predicts downstream failure in real synthetic-data training loops for modern LLMs and changes concrete training decisions at scale.",
      "one_sentence_episode_hook": "If future LLMs are trained on AI-made text, can we detect the exact spectral moment when their world starts to collapse?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02970",
      "title": "Reliability-Aware Adaptive Self-Consistency for Efficient Sampling in LLM Reasoning",
      "abstract": "Self-Consistency improves reasoning reliability through multi-sample aggregation, but incurs substantial inference cost. Adaptive self-consistency methods mitigate this issue by adjusting the sampling budget; however, they rely on count-based stopping rules that treat all responses equally, often leading to unnecessary sampling. We propose Reliability-Aware Adaptive Self-Consistency (ReASC), which addresses this limitation by reframing adaptive sampling from response counting to evidence sufficiency, leveraging response-level confidence for principled information aggregation. ReASC operates in two stages: a single-sample decision stage that resolves instances confidently answerable from a single response, and a reliability-aware accumulation stage that aggregates responses by jointly leveraging their frequency and confidence. Across five models and four datasets, ReASC consistently achieves the best accuracy-cost trade-off compared to existing baselines, yielding improved inference efficiency across model scales from 3B to 27B parameters. As a concrete example, ReASC reduces inference cost by up to 70\\% relative to self-consistency while preserving accuracy on GSM8K using Gemma-3-4B-it.",
      "authors": [
        "Junseok Kim",
        "Nakyeong Yang",
        "Kyungmin Min",
        "Kyomin Jung"
      ],
      "published_at": "2026-01-06T12:27:53",
      "categories": [
        "cs.CL",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.02970v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.3496623933315277,
      "sim_memory": 0.34933343529701233,
      "sim_negative": 0.17208585143089294,
      "broad_relevance": 0.3496623933315277,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5445442795753479,
      "evidence_score": 0.5493147180559945,
      "direct_memory_relevance": 0.34933343529701233,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6699999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5204757578370596,
      "memory_score": 0.11611779675582304,
      "quality_score": 0.5727258872223978,
      "bridge_score": 0.11611779675582304,
      "max_axis_score": 0.5204757578370596,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference",
        "Systems"
      ],
      "status": "Monitor",
      "why_now": "Inference cost for reasoning-heavy LLM use is a live issue, and this paper targets a practical pain point: reducing self-consistency sampling without giving up accuracy. The proposed reliability-aware stopping rule is broadly relevant to anyone deploying multi-sample reasoning pipelines.",
      "why_not_higher": "The contribution is mainly an inference-time aggregation heuristic for reasoning benchmarks rather than a deeper shift in model design or serving architecture. Its memory/storage relevance is indirect, coming through reduced sampling cost rather than any explicit cache, bandwidth, or offload mechanism.",
      "downgrade_reasons": [
        "memory connection is only indirect",
        "benchmark-centered reasoning paper",
        "limited evidence of impact on real production serving stacks",
        "incremental over prior adaptive self-consistency methods"
      ],
      "what_would_raise_priority": "Show robust gains across harder real-world agent or long-context workloads and connect the method to end-to-end serving efficiency under realistic latency and throughput constraints.",
      "one_sentence_episode_hook": "Can confidence-aware stopping rules make chain-of-thought voting dramatically cheaper without sacrificing the reliability gains that made self-consistency popular in the first place?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03066",
      "title": "Do LLMs Encode Functional Importance of Reasoning Tokens?",
      "abstract": "Large language models solve complex tasks by generating long reasoning chains, achieving higher accuracy at the cost of increased computational cost and reduced ability to isolate functionally relevant reasoning. Prior work on compact reasoning shortens such chains through probabilistic sampling, heuristics, or supervision from frontier models, but offers limited insight into whether models internally encode token-level functional importance for answer generation. We address this gap diagnostically and propose greedy pruning, a likelihood-preserving deletion procedure that iteratively removes reasoning tokens whose removal minimally degrades model likelihood under a specified objective, yielding length-controlled reasoning chains. We evaluate pruned reasoning in a distillation framework and show that students trained on pruned chains outperform a frontier-model-supervised compression baseline at matched reasoning lengths. Finally, our analysis reveals systematic pruning patterns and shows that attention scores can predict greedy pruning ranks, further suggesting that models encode a nontrivial functional importance structure over reasoning tokens.",
      "authors": [
        "Janvijay Singh",
        "Dilek Hakkani-T\u00fcr"
      ],
      "published_at": "2026-01-06T14:50:02",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03066v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 2,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.467390775680542,
      "sim_memory": 0.375352144241333,
      "sim_negative": 0.19458933174610138,
      "broad_relevance": 0.467390775680542,
      "momentum": 0.031333575283049045,
      "teachability": 0.75,
      "novelty_score": 0.5784153938293457,
      "evidence_score": 0.529861228866811,
      "direct_memory_relevance": 0.375352144241333,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8599999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.549725441165196,
      "memory_score": 0.08578899073641721,
      "quality_score": 0.6189444915467244,
      "bridge_score": 0.08578899073641721,
      "max_axis_score": 0.549725441165196,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "Reasoning-token pruning is timely because the field is actively trying to make chain-of-thought cheaper, shorter, and more interpretable without losing accuracy. The paper also asks a broadly interesting question: whether LLMs internally encode which reasoning tokens actually matter for the answer.",
      "why_not_higher": "The contribution is more diagnostic and distillation-oriented than directly practice-changing for deployed systems. Its memory/storage relevance is weak because it is about reasoning compression rather than concrete KV-cache, offload, paging, bandwidth, or scheduling mechanisms.",
      "downgrade_reasons": [
        "memory connection is indirect",
        "limited direct systems impact",
        "reasoning/distillation space is crowded",
        "unclear validation on frontier-scale production workloads"
      ],
      "what_would_raise_priority": "Show consistent gains on widely used frontier reasoning models with clear latency, token-cost, or serving benefits at deployment scale.",
      "one_sentence_episode_hook": "Can an LLM tell which parts of its own chain-of-thought actually matter\u2014and can we delete the rest without losing the answer?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03401",
      "title": "Rendering Data Unlearnable by Exploiting LLM Alignment Mechanisms",
      "abstract": "Large language models (LLMs) are increasingly trained on massive, heterogeneous text corpora, raising serious concerns about the unauthorised use of proprietary or personal data during model training. In this work, we address the problem of data protection against unwanted model learning in a realistic black-box setting. We propose Disclaimer Injection, a novel data-level defence that renders text unlearnable to LLMs. Rather than relying on model-side controls or explicit data removal, our approach exploits the models' own alignment mechanisms: by injecting carefully designed alignment-triggering disclaimers to prevent effective learning. Through layer-wise analysis, we find that fine-tuning on such protected data induces persistent activation of alignment-related layers, causing alignment constraints to override task learning even on common inputs. Consequently, models trained on such data exhibit substantial and systematic performance degradation compared to standard fine-tuning. Our results identify alignment behaviour as a previously unexplored lever for data protection and, to our knowledge, present the first practical method for restricting data learnability at LLM scale without requiring access to or modification of the training pipeline.",
      "authors": [
        "Ruihan Zhang",
        "Jun Sun"
      ],
      "published_at": "2026-01-06T20:34:15",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03401v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.45104512572288513,
      "sim_memory": 0.40529462695121765,
      "sim_negative": 0.20963937044143677,
      "broad_relevance": 0.45104512572288513,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5448887944221497,
      "evidence_score": 0.32,
      "direct_memory_relevance": 0.40529462695121765,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.64,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.567046856880188,
      "memory_score": 0.05158838808536528,
      "quality_score": 0.545,
      "bridge_score": 0.05158838808536528,
      "max_axis_score": 0.567046856880188,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "Data provenance, consent, and training-time protection are timely concerns as LLM developers continue ingesting large web-scale corpora. A black-box method that tries to make text 'unlearnable' by exploiting alignment behavior is novel enough to attract broad AI curiosity.",
      "why_not_higher": "This is mainly a data-protection/alignment intervention rather than a memory or systems paper, and the abstract emphasizes performance degradation under fine-tuning more than robust validation on diverse real-world training setups. Its practical reliability against varied pipelines, data cleaning, and stronger training regimes is still unclear.",
      "downgrade_reasons": [
        "Not a memory/storage paper",
        "Evidence appears limited from abstract",
        "May be brittle to preprocessing or training changes",
        "More of a defensive alignment/training curiosity than a deployment-changing result"
      ],
      "what_would_raise_priority": "Show consistent effects across multiple frontier-relevant models, training pipelines, and realistic corpus-cleaning settings, with clear failure modes and practical deployment guidance.",
      "one_sentence_episode_hook": "What if the best way to stop an LLM from learning your text is to weaponize its own alignment training against it?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02978",
      "title": "Mechanistic Knobs in LLMs: Retrieving and Steering High-Order Semantic Features via Sparse Autoencoders",
      "abstract": "Recent work in Mechanistic Interpretability (MI) has enabled the identification and intervention of internal features in Large Language Models (LLMs). However, a persistent challenge lies in linking such internal features to the reliable control of complex, behavior-level semantic attributes in language generation. In this paper, we propose a Sparse Autoencoder-based framework for retrieving and steering semantically interpretable internal features associated with high-level linguistic behaviors. Our method employs a contrastive feature retrieval pipeline based on controlled semantic oppositions, combing statistical activation analysis and generation-based validation to distill monosemantic functional features from sparse activation spaces. Using the Big Five personality traits as a case study, we demonstrate that our method enables precise, bidirectional steering of model behavior while maintaining superior stability and performance compared to existing activation steering methods like Contrastive Activation Addition (CAA). We further identify an empirical effect, which we term Functional Faithfulness, whereby intervening on a specific internal feature induces coherent and predictable shifts across multiple linguistic dimensions aligned with the target semantic attribute. Our findings suggest that LLMs internalize deeply integrated representations of high-order concepts, and provide a novel, robust mechanistic path for the regulation of complex AI behaviors.",
      "authors": [
        "Ruikang Zhang",
        "Shuo Wang",
        "Qi Su"
      ],
      "published_at": "2026-01-06T12:40:37",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.02978v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.4313969910144806,
      "sim_memory": 0.40109413862228394,
      "sim_negative": 0.22054529190063477,
      "broad_relevance": 0.4313969910144806,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.58905029296875,
      "evidence_score": 0.24,
      "direct_memory_relevance": 0.40109413862228394,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.8799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.3,
      "public_interest_score": 0.5427766412496566,
      "memory_score": 0.03132824158668518,
      "quality_score": 0.525,
      "bridge_score": 0.03132824158668518,
      "max_axis_score": 0.5427766412496566,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Theory",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "Mechanistic interpretability remains a live topic, and this paper pushes from identifying features toward actually steering higher-level behaviors like personality traits in generation. That makes it timely for listeners interested in controllable LLMs and whether internal representations can become usable control knobs.",
      "why_not_higher": "The work is not meaningfully about memory/storage systems, and the evidence in the abstract appears centered on a case study rather than broad validation across frontier models, tasks, and deployment settings. It sounds interesting editorially, but not yet like a clear practice-changing result.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Case-study framing around Big Five traits",
        "Unclear realism of workloads and model scale",
        "Limited evidence of impact on training/serving pipelines",
        "Interpretability-steering space is crowded and easy to overclaim"
      ],
      "what_would_raise_priority": "Show robust transfer across multiple strong open and closed LLMs, broader semantic attributes, and real downstream evaluation compared against standard steering baselines.",
      "one_sentence_episode_hook": "Can sparse autoencoders turn hidden LLM features into reliable knobs for steering personality-level behavior, or is this another neat interpretability demo that breaks outside the lab?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02997",
      "title": "From Memorization to Creativity: LLM as a Designer of Novel Neural-Architectures",
      "abstract": "Large language models (LLMs) excel in program synthesis, yet their ability to autonomously navigate neural architecture design--balancing syntactic reliability, performance, and structural novelty--remains underexplored. We address this by placing a code-oriented LLM within a closed-loop synthesis framework, analyzing its evolution over 22 supervised fine-tuning cycles. The model synthesizes PyTorch convolutional networks which are validated, evaluated via low-fidelity performance signals (single-epoch accuracy), and filtered using a MinHash-Jaccard criterion to prevent structural redundancy. High-performing, novel architectures are converted into prompt-code pairs for iterative fine-tuning via parameter-efficient LoRA adaptation, initialized from the LEMUR dataset. Across cycles, the LLM internalizes empirical architectural priors, becoming a robust generator. The valid generation rate stabilizes at 50.6 percent (peaking at 74.5 percent), while mean first-epoch accuracy rises from 28.06 percent to 50.99 percent, and the fraction of candidates exceeding 40 percent accuracy grows from 2.04 percent to 96.81 percent. Analyses confirm the model moves beyond replicating existing motifs, synthesizing 455 high-performing architectures absent from the original corpus. By grounding code synthesis in execution feedback, this work provides a scalable blueprint for transforming stochastic generators into autonomous, performance-driven neural designers, establishing that LLMs can internalize empirical, non-textual rewards to transcend their training data.",
      "authors": [
        "Waleed Khalid",
        "Dmitry Ignatov",
        "Radu Timofte"
      ],
      "published_at": "2026-01-06T13:20:28",
      "categories": [
        "cs.LG",
        "cs.CV"
      ],
      "url": "http://arxiv.org/abs/2601.02997v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 3,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "vision",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5264967679977417,
      "sim_memory": 0.43937206268310547,
      "sim_negative": 0.2973763346672058,
      "broad_relevance": 0.5264967679977417,
      "momentum": 0.04279025907815965,
      "teachability": 0.75,
      "novelty_score": 0.4276317358016968,
      "evidence_score": 0.458629436111989,
      "direct_memory_relevance": 0.43937206268310547,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.54,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.49144625800200736,
      "memory_score": 0.024447151138370288,
      "quality_score": 0.5804517744447956,
      "bridge_score": 0.024447151138370288,
      "max_axis_score": 0.49144625800200736,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training"
      ],
      "status": "Monitor",
      "why_now": "LLM-driven neural architecture search is a timely question because labs are actively probing whether code models can move from autocomplete to closed-loop scientific design. The paper offers a concrete iterative recipe\u2014generation, execution feedback, novelty filtering, and LoRA updates\u2014that could interest a broad AI audience.",
      "why_not_higher": "The demonstrated setting is narrow: PyTorch CNN design for vision with low-fidelity single-epoch signals, not modern frontier model training or serving. Its connection to the podcast's memory/storage lens is essentially absent, and the evidence for broad transfer beyond this task is limited.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "CNN architecture search is narrower than current transformer-centric practice",
        "Uses low-fidelity proxy evaluation rather than realistic end-to-end workloads",
        "Unclear transfer to modern large-model build or serve pipelines"
      ],
      "what_would_raise_priority": "Show the same closed-loop method producing useful, novel improvements for transformer architectures or training/serving systems under realistic compute constraints.",
      "one_sentence_episode_hook": "Can an LLM learn from execution feedback to invent genuinely new neural networks\u2014or is this still just smarter remixing of old architecture motifs?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03193",
      "title": "UniCorn: Towards Self-Improving Unified Multimodal Models through Self-Generated Supervision",
      "abstract": "While Unified Multimodal Models (UMMs) have achieved remarkable success in cross-modal comprehension, a significant gap persists in their ability to leverage such internal knowledge for high-quality generation. We formalize this discrepancy as Conduction Aphasia, a phenomenon where models accurately interpret multimodal inputs but struggle to translate that understanding into faithful and controllable synthesis. To address this, we propose UniCorn, a simple yet elegant self-improvement framework that eliminates the need for external data or teacher supervision. By partitioning a single UMM into three collaborative roles: Proposer, Solver, and Judge, UniCorn generates high-quality interactions via self-play and employs cognitive pattern reconstruction to distill latent understanding into explicit generative signals. To validate the restoration of multimodal coherence, we introduce UniCycle, a cycle-consistency benchmark based on a Text to Image to Text reconstruction loop. Extensive experiments demonstrate that UniCorn achieves comprehensive and substantial improvements over the base model across six general image generation benchmarks. Notably, it achieves SOTA performance on TIIF(73.8), DPG(86.8), CompBench(88.5), and UniCycle while further delivering substantial gains of +5.0 on WISE and +6.5 on OneIG. These results highlight that our method significantly enhances T2I generation while maintaining robust comprehension, demonstrating the scalability of fully self-supervised refinement for unified multimodal intelligence.",
      "authors": [
        "Ruiyan Han",
        "Zhen Fang",
        "XinYu Sun",
        "Yuchen Ma",
        "Ziheng Wang",
        "Yu Zeng",
        "Zehui Chen",
        "Lin Chen",
        "Wenxuan Huang",
        "Wei-Jie Xu"
      ],
      "published_at": "2026-01-06T17:15:50",
      "categories": [
        "cs.CV",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03193v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 5,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "vision",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4639514982700348,
      "sim_memory": 0.3986424207687378,
      "sim_negative": 0.28719496726989746,
      "broad_relevance": 0.4639514982700348,
      "momentum": 0.06108604879161034,
      "teachability": 0.7,
      "novelty_score": 0.5062346458435059,
      "evidence_score": 0.5591759469228055,
      "direct_memory_relevance": 0.3986424207687378,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.64,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.5342142481542792,
      "memory_score": 0.05709383986135799,
      "quality_score": 0.6166703787691222,
      "bridge_score": 0.05709383986135799,
      "max_axis_score": 0.5342142481542792,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "Self-improvement without external teachers is a timely theme, and this paper targets a real weakness in unified multimodal models: they can understand inputs better than they can faithfully generate from that understanding. The proposer-solver-judge setup and the new cycle-consistency benchmark make it potentially interesting to a broad multimodal audience.",
      "why_not_higher": "The contribution is mostly a training recipe plus benchmark gains for text-to-image generation within unified multimodal models, not a clear shift in how mainstream AI systems are built or served. It also has little direct connection to memory, storage, or systems behavior central to the podcast\u2019s second lens.",
      "downgrade_reasons": [
        "Weak memory/storage relevance",
        "Primarily benchmark and training-framework driven",
        "Unclear adoption path beyond this model family",
        "Evidence appears concentrated on benchmark suites rather than deployment realism"
      ],
      "what_would_raise_priority": "Stronger evidence that the self-improvement method generalizes across multiple open multimodal model families and changes practical training pipelines would raise priority.",
      "one_sentence_episode_hook": "Can a multimodal model teach itself to turn what it already understands into better images\u2014without new data, teachers, or synthetic labels from a stronger model?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03137",
      "title": "Accurate Table Question Answering with Accessible LLMs",
      "abstract": "Given a table T in a database and a question Q in natural language, the table question answering (TQA) task aims to return an accurate answer to Q based on the content of T. Recent state-of-the-art solutions leverage large language models (LLMs) to obtain high-quality answers. However, most rely on proprietary, large-scale LLMs with costly API access, posing a significant financial barrier. This paper instead focuses on TQA with smaller, open-weight LLMs that can run on a desktop or laptop. This setting is challenging, as such LLMs typically have weaker capabilities than large proprietary models, leading to substantial performance degradation with existing methods.   We observe that a key reason for this degradation is that prior approaches often require the LLM to solve a highly sophisticated task using long, complex prompts, which exceed the capabilities of small open-weight LLMs. Motivated by this observation, we present Orchestra, a multi-agent approach that unlocks the potential of accessible LLMs for high-quality, cost-effective TQA. Orchestra coordinates a group of LLM agents, each responsible for a relatively simple task, through a structured, layered workflow to solve complex TQA problems -- akin to an orchestra. By reducing the prompt complexity faced by each agent, Orchestra significantly improves output reliability.   We implement Orchestra on top of AgentScope, an open-source multi-agent framework, and evaluate it on multiple TQA benchmarks using a wide range of open-weight LLMs. Experimental results show that Orchestra achieves strong performance even with small- to medium-sized models. For example, with Qwen2.5-14B, Orchestra reaches 72.1% accuracy on WikiTQ, approaching the best prior result of 75.3% achieved with GPT-4; with larger Qwen, Llama, or DeepSeek models, Orchestra outperforms all prior methods and establishes new state-of-the-art results across all benchmarks.",
      "authors": [
        "Yangfan Jiang",
        "Fei Wei",
        "Ergute Bao",
        "Yaliang Li",
        "Bolin Ding",
        "Yin Yang",
        "Xiaokui Xiao"
      ],
      "published_at": "2026-01-06T16:07:25",
      "categories": [
        "cs.DB",
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03137v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.5030900835990906,
      "sim_memory": 0.37993746995925903,
      "sim_negative": 0.18813201785087585,
      "broad_relevance": 0.5030900835990906,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5169882774353027,
      "evidence_score": 0.46,
      "direct_memory_relevance": 0.37993746995925903,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5184752666950225,
      "memory_score": 0.10498124098777767,
      "quality_score": 0.575,
      "bridge_score": 0.10498124098777767,
      "max_axis_score": 0.5184752666950225,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Systems",
        "Inference",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "It speaks to a timely and practical question: how far smaller open-weight LLMs can go on real tasks when the workflow is redesigned instead of just scaling model size. The cost-accessibility angle and strong benchmark results make it relevant to many builders evaluating open models versus proprietary APIs.",
      "why_not_higher": "This is still primarily a table-QA workflow paper rather than a broadly general method for modern AI systems, and the multi-agent decomposition is not clearly transformative beyond this task family. It also has little direct connection to memory, storage, or serving mechanics.",
      "downgrade_reasons": [
        "Narrow task focus on table question answering",
        "Memory/storage link is absent rather than central",
        "Multi-agent orchestration may be prompt-engineering-heavy and task-specific",
        "Unclear how well the gains transfer beyond TQA benchmarks"
      ],
      "what_would_raise_priority": "Evidence that the structured decomposition consistently improves other enterprise data tasks or agentic retrieval settings with small open models would make it more broadly compelling.",
      "one_sentence_episode_hook": "Can clever orchestration let a laptop-friendly open model nearly catch GPT-4 on table QA without changing the model at all?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03396",
      "title": "Breaking the Assistant Mold: Modeling Behavioral Variation in LLM Based Procedural Character Generation",
      "abstract": "Procedural content generation has enabled vast virtual worlds through levels, maps, and quests, but large-scale character generation remains underexplored. We identify two alignment-induced biases in existing methods: a positive moral bias, where characters uniformly adopt agreeable stances (e.g. always saying lying is bad), and a helpful assistant bias, where characters invariably answer questions directly (e.g. never refusing or deflecting). While such tendencies suit instruction-following systems, they suppress dramatic tension and yield predictable characters, stemming from maximum likelihood training and assistant fine-tuning. To address this, we introduce PersonaWeaver, a framework that disentangles world-building (roles, demographics) from behavioral-building (moral stances, interactional styles), yielding characters with more diverse reactions and moral stances, as well as second-order diversity in stylistic markers like length, tone, and punctuation. Code: https://github.com/mqraitem/Persona-Weaver",
      "authors": [
        "Maan Qraitem",
        "Kate Saenko",
        "Bryan A. Plummer"
      ],
      "published_at": "2026-01-06T20:18:01",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03396v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4305998980998993,
      "sim_memory": 0.29086804389953613,
      "sim_negative": 0.1781441867351532,
      "broad_relevance": 0.4305998980998993,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6314728558063507,
      "evidence_score": 0.36000000000000004,
      "direct_memory_relevance": 0.29086804389953613,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.2,
      "transferability_score": 0.82,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5139008978009224,
      "memory_score": 0.020260413169860825,
      "quality_score": 0.575,
      "bridge_score": 0.020260413169860825,
      "max_axis_score": 0.5139008978009224,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "As LLMs get used for games, agents, and social simulation, the default helpful-assistant behavior is becoming a visible limitation. A framework that explicitly models behavioral and moral variation taps into a timely conversation about alignment side effects and controllable generation.",
      "why_not_higher": "This is not a memory/storage paper, and its impact on mainstream model building or serving is indirect. It also appears focused on procedural character generation rather than a broadly adopted systems or evaluation change for modern LLM pipelines.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Application-leaning character generation focus",
        "Limited evidence of effect on mainstream training or serving practice",
        "Likely evaluation centered on generated personas rather than realistic deployment workloads"
      ],
      "what_would_raise_priority": "Stronger evidence that the method improves broadly useful persona control and evaluation across common LLM products, agent benchmarks, or game-scale deployments would raise priority.",
      "one_sentence_episode_hook": "What if alignment made every AI character too nice and too helpful\u2014and how do you break that assistant mold without breaking the model?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03134",
      "title": "The Anatomy of Conversational Scams: A Topic-Based Red Teaming Analysis of Multi-Turn Interactions in LLMs",
      "abstract": "As LLMs gain persuasive agentic capabilities through extended dialogues, they introduce novel risks in multi-turn conversational scams that single-turn safety evaluations fail to capture. We systematically study these risks using a controlled LLM-to-LLM simulation framework across multi-turn scam scenarios. Evaluating eight state-of-the-art models in English and Chinese, we analyze dialogue outcomes and qualitatively annotate attacker strategies, defensive responses, and failure modes. Results reveal that scam interactions follow recurrent escalation patterns, while defenses employ verification and delay mechanisms. Furthermore, interactional failures frequently stem from safety guardrail activation and role instability. Our findings highlight multi-turn interactional safety as a critical, distinct dimension of LLM behavior.",
      "authors": [
        "Xiangzhe Yuan",
        "Zhenhao Zhang",
        "Haoming Tang",
        "Siying Hu"
      ],
      "published_at": "2026-01-06T16:06:04",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03134v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4136293828487396,
      "sim_memory": 0.12546895444393158,
      "sim_negative": 0.15339720249176025,
      "broad_relevance": 0.4136293828487396,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6651434898376465,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.12546895444393158,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.82,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.5338603383302689,
      "memory_score": 0.0006406863331794577,
      "quality_score": 0.575,
      "bridge_score": 0.0006406863331794577,
      "max_axis_score": 0.5338603383302689,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Application"
      ],
      "status": "Monitor",
      "why_now": "As LLMs are increasingly deployed in long-running assistant and agent settings, multi-turn scam susceptibility is a timely safety issue that single-turn red teaming misses. The paper also looks at multiple frontier models across two languages, which helps it land beyond a toy case study.",
      "why_not_higher": "It is primarily a conversational safety analysis rather than a paper that changes core model building, serving, or systems practice. The memory/storage angle is absent, and the evaluation relies on simulated LLM-to-LLM interactions rather than real user deployments.",
      "downgrade_reasons": [
        "No direct memory/storage contribution",
        "More safety benchmarking than systems or training impact",
        "Simulation-based evidence rather than real-world deployment data",
        "Narrower editorial fit than core AI infrastructure papers"
      ],
      "what_would_raise_priority": "Stronger evidence that the identified failure modes alter mainstream evaluation protocols or deployment guardrails for production LLM agents would raise its priority.",
      "one_sentence_episode_hook": "If LLM safety tests are mostly single-turn, are we missing the real danger zone where scams slowly win over a model across a whole conversation?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.04254",
      "title": "Scaling Trends for Multi-Hop Contextual Reasoning in Mid-Scale Language Models",
      "abstract": "We present a controlled study of multi-hop contextual reasoning in large language models, providing a clean demonstration of the task-method dissociation: rule-based pattern matching achieves 100% success on structured information retrieval but only 6.7% on tasks requiring cross-document reasoning, while LLM-based multi-agent systems show the inverse pattern, achieving up to 80% on reasoning tasks where rule-based methods fail. Using a synthetic evaluation framework with 120 trials across four models (LLaMA-3 8B, LLaMA-2 13B, Mixtral 8x7B, DeepSeek-V2 16B), we report three key findings: (1) Multi-agent amplification depends on base capability: statistically significant gains occur only for models with sufficient reasoning ability (p < 0.001 for LLaMA-3 8B, p = 0.014 for Mixtral), with improvements of up to 46.7 percentage points, while weaker models show no benefit, suggesting amplification rather than compensation; (2) Active parameters predict reasoning performance: Mixtral's performance aligns with its ~12B active parameters rather than 47B total, consistent with the hypothesis that inference-time compute drives reasoning capability in MoE architectures; (3) Architecture quality matters: LLaMA-3 8B outperforms LLaMA-2 13B despite fewer parameters, consistent with known training improvements. Our results provide controlled quantitative evidence for intuitions about multi-agent coordination and MoE scaling, while highlighting the dependence of multi-agent benefits on base model capability. We release our evaluation framework to support reproducible research on reasoning in mid-scale models.",
      "authors": [
        "Brady Steele",
        "Micah Katz"
      ],
      "published_at": "2026-01-06T20:18:55",
      "categories": [
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.04254v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5077535510063171,
      "sim_memory": 0.36607614159584045,
      "sim_negative": 0.21512022614479065,
      "broad_relevance": 0.5077535510063171,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.48468661308288574,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.36607614159584045,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7599999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.515029057264328,
      "memory_score": 0.05282284247875213,
      "quality_score": 0.575,
      "bridge_score": 0.05282284247875213,
      "max_axis_score": 0.515029057264328,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "There is active interest in whether multi-agent orchestration actually improves reasoning, and this paper offers a clean controlled result: it helps only when the base model is already capable. That makes it timely for practitioners evaluating agentic scaffolds around mid-scale open models.",
      "why_not_higher": "The study is synthetic, small-scale, and centered on a narrow contextual reasoning benchmark rather than realistic production workloads. It also has no meaningful memory/storage mechanism despite some compute-parameter discussion around MoE active parameters.",
      "downgrade_reasons": [
        "not a memory/storage paper",
        "synthetic benchmark with 120 trials",
        "limited model set and mid-scale focus",
        "weak evidence of deployment impact",
        "more evaluation result than systems change"
      ],
      "what_would_raise_priority": "Show the same capability-threshold effect on broader agent benchmarks or real retrieval/workflow tasks, ideally with implications for inference cost and serving design.",
      "one_sentence_episode_hook": "Do multi-agent LLM systems really make weak models smarter, or do they only amplify models that could already reason in the first place?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03052",
      "title": "Detecting Hallucinations in Retrieval-Augmented Generation via Semantic-level Internal Reasoning Graph",
      "abstract": "The Retrieval-augmented generation (RAG) system based on Large language model (LLM) has made significant progress. It can effectively reduce factuality hallucinations, but faithfulness hallucinations still exist. Previous methods for detecting faithfulness hallucinations either neglect to capture the models' internal reasoning processes or handle those features coarsely, making it difficult for discriminators to learn. This paper proposes a semantic-level internal reasoning graph-based method for detecting faithfulness hallucination. Specifically, we first extend the layer-wise relevance propagation algorithm from the token level to the semantic level, constructing an internal reasoning graph based on attribution vectors. This provides a more faithful semantic-level representation of dependency. Furthermore, we design a general framework based on a small pre-trained language model to utilize the dependencies in LLM's reasoning for training and hallucination detection, which can dynamically adjust the pass rate of correct samples through a threshold. Experimental results demonstrate that our method achieves better overall performance compared to state-of-the-art baselines on RAGTruth and Dolly-15k.",
      "authors": [
        "Jianpeng Hu",
        "Yanzeng Li",
        "Jialun Zhong",
        "Wenfa Qi",
        "Lei Zou"
      ],
      "published_at": "2026-01-06T14:35:20",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03052v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4315766990184784,
      "sim_memory": 0.3210328221321106,
      "sim_negative": 0.2769216299057007,
      "broad_relevance": 0.4315766990184784,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5669727325439453,
      "evidence_score": 0.36000000000000004,
      "direct_memory_relevance": 0.3210328221321106,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.74,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.44451891958713535,
      "memory_score": 0.019309846639633144,
      "quality_score": 0.575,
      "bridge_score": 0.019309846639633144,
      "max_axis_score": 0.44451891958713535,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference"
      ],
      "status": "Monitor",
      "why_now": "RAG hallucination detection remains a live topic, and this paper targets the harder faithfulness problem by trying to read internal reasoning structure rather than only comparing outputs to retrieved context. That makes it timely for teams building evaluation and guardrail layers around LLM applications.",
      "why_not_higher": "The contribution looks like a specialized hallucination-detection method rather than a change to core model building or serving, and the memory/storage link is essentially absent. Evidence is benchmark-centric and does not yet show broad operational impact across production-scale RAG settings or major frontier models.",
      "downgrade_reasons": [
        "memory/storage connection absent",
        "mostly evaluation/guardrail rather than core systems change",
        "benchmark result on narrow task framing",
        "unclear robustness on production RAG pipelines and top-tier models"
      ],
      "what_would_raise_priority": "Stronger cross-model, production-style RAG evidence showing this detector materially improves deployed system reliability or evaluation practice would raise priority.",
      "one_sentence_episode_hook": "Can you catch RAG hallucinations by peeking inside an LLM's own reasoning graph instead of just fact-checking the final answer?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "deferred": [
    {
      "arxiv_id": "2601.03089",
      "title": "Grad-ELLM: Gradient-based Explanations for Decoder-only LLMs",
      "abstract": "Large Language Models (LLMs) have demonstrated remarkable capabilities across diverse tasks, yet their black-box nature raises concerns about transparency and faithfulness. Input attribution methods aim to highlight each input token's contributions to the model's output, but existing approaches are typically model-agnostic, and do not focus on transformer-specific architectures, leading to limited faithfulness. To address this, we propose Grad-ELLM, a gradient-based attribution method for decoder-only transformer-based LLMs. By aggregating channel importance from gradients of the output logit with respect to attention layers and spatial importance from attention maps, Grad-ELLM generates heatmaps at each generation step without requiring architectural modifications. Additionally, we introduce two faithfulneses metrics $\u03c0$-Soft-NC and $\u03c0$-Soft-NS, which are modifications of Soft-NC/NS that provide fairer comparisons by controlling the amount of information kept when perturbing the text. We evaluate Grad-ELLM on sentiment classification, question answering, and open-generation tasks using different models. Experiment results show that Grad-ELLM consistently achieves superior faithfulness than other attribution methods.",
      "authors": [
        "Xin Huang",
        "Antoni B. Chan"
      ],
      "published_at": "2026-01-06T15:22:39",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03089v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.46640658378601074,
      "sim_memory": 0.4724682867527008,
      "sim_negative": 0.28225284814834595,
      "broad_relevance": 0.46640658378601074,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.4773273468017578,
      "evidence_score": 0.37,
      "direct_memory_relevance": 0.4724682867527008,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.82,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.48152107715606685,
      "memory_score": 0.017740486025810215,
      "quality_score": 0.595,
      "bridge_score": 0.017740486025810215,
      "max_axis_score": 0.48152107715606685,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "As LLMs are deployed more widely, faithfulness of explanations and token attribution remains a live concern for evaluation and trust. A decoder-only-specific attribution method is timely because most practical LLM use is in autoregressive inference.",
      "why_not_higher": "This is mainly an interpretability/evaluation paper, not something that clearly changes how most teams build or serve models in the near term. The memory/storage angle is essentially absent, and the practical impact depends on whether attribution methods become part of standard LLM eval stacks.",
      "downgrade_reasons": [
        "No direct memory/storage contribution",
        "Primarily interpretability rather than systems or deployment",
        "Unclear adoption beyond attribution benchmarks",
        "May be incremental within gradient-based explanation literature"
      ],
      "what_would_raise_priority": "Evidence that Grad-ELLM becomes a practical debugging or safety-eval tool on frontier-scale LLM workflows would raise its editorial priority.",
      "one_sentence_episode_hook": "Can we make decoder-only LLM explanations more faithful by using the model's own gradients and attention, instead of treating transformers like generic black boxes?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03079",
      "title": "Enhancing Moral Diagnosis and Correction in Large Language Models",
      "abstract": "Identifying specific moral errors in an input and generating appropriate corrections require moral sensitivity in large language models (LLMs), which is fundamental for developing their moral performance, yet a challenging task. This study leverages a pragmatic inference-based approach to enhance both the moral diagnosis and corrections of models. Crucially, our method generalizes across a diverse set of different tasks, including moral reasoning, toxic language detection, social bias detection, and jailbreaks, despite substantial differences in their semantic formulations. To enable such generalization, the study also introduces a unifying variable, pragmatic inference load, which captures the degree of pragmatic reasoning required across tasks. Experimental results show that our approach enables LLMs to produce high-quality diagnostic outputs of moral errors, make effective corrections, and consistently outperform a range of baseline methods. Further analyses reveal that these improvements do not arise from heuristic-based response patterns, but from learned inferential processes, highlighting the effectiveness of our approach.",
      "authors": [
        "Bocheng Chen",
        "Xi Chen",
        "Han Zi",
        "Haitao Mao",
        "Zimo Qi",
        "Xitong Zhang",
        "Kristen Johnson",
        "Guangliang Liu"
      ],
      "published_at": "2026-01-06T15:09:05",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03079v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4792913794517517,
      "sim_memory": 0.3377918004989624,
      "sim_negative": 0.1634984165430069,
      "broad_relevance": 0.4792913794517517,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6295172572135925,
      "evidence_score": 0.36000000000000004,
      "direct_memory_relevance": 0.3377918004989624,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.83,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.4282150024175644,
      "memory_score": 0.0,
      "quality_score": 0.575,
      "bridge_score": 0.0,
      "max_axis_score": 0.4282150024175644,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "LLM behavior on moral diagnosis, toxicity, bias, and jailbreak correction is timely as labs look for better post-training and evaluation methods for safer deployment. The paper also tries to unify these tasks through a pragmatic-inference framing rather than treating each as a separate benchmark.",
      "why_not_higher": "This is not a memory/storage paper, and for a broad AI audience it looks more like a safety-behavior benchmarking and prompting/evaluation method than a change to mainstream model building or serving. The abstract claims broad generalization, but without clear evidence here on frontier-scale models or production-like deployment settings, the practical impact is uncertain.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Primarily safety/alignment behavior paper",
        "Unclear systems impact on training or serving",
        "Evidence on realistic large-scale deployments not clear from abstract",
        "Could become a benchmark-centric episode rather than a broadly educational one"
      ],
      "what_would_raise_priority": "Stronger evidence that the method reliably improves frontier LLM safety behavior in production-style evals and transfers into deployable post-training or inference pipelines would raise it.",
      "one_sentence_episode_hook": "Can we teach LLMs not just to spot moral failures, but to explain exactly what went wrong and fix it across toxicity, bias, and jailbreaks using one pragmatic-inference lens?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03359",
      "title": "Enhancing LLM Instruction Following: An Evaluation-Driven Multi-Agentic Workflow for Prompt Instructions Optimization",
      "abstract": "Large Language Models (LLMs) often generate substantively relevant content but fail to adhere to formal constraints, leading to outputs that are conceptually correct but procedurally flawed. Traditional prompt refinement approaches focus on rephrasing the description of the primary task an LLM has to perform, neglecting the granular constraints that function as acceptance criteria for its response. We propose a novel multi-agentic workflow that decouples optimization of the primary task description from its constraints, using quantitative scores as feedback to iteratively rewrite and improve them. Our evaluation demonstrates this method produces revised prompts that yield significantly higher compliance scores from models like Llama 3.1 8B and Mixtral-8x 7B.",
      "authors": [
        "Alberto Purpura",
        "Li Wang",
        "Sahil Badyal",
        "Eugenio Beaufrand",
        "Adam Faulkner"
      ],
      "published_at": "2026-01-06T19:02:14",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03359v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.5198076963424683,
      "sim_memory": 0.2874417304992676,
      "sim_negative": 0.17588120698928833,
      "broad_relevance": 0.5198076963424683,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.48709380626678467,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.2874417304992676,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.58,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.4190063798427582,
      "memory_score": 0.0022325191497802876,
      "quality_score": 0.545,
      "bridge_score": 0.0022325191497802876,
      "max_axis_score": 0.4190063798427582,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "Instruction following remains a practical pain point for teams deploying open LLMs, and the paper frames prompt optimization around measurable constraint compliance rather than vague prompt tinkering. That makes it timely as practitioners look for more systematic ways to improve output reliability without retraining models.",
      "why_not_higher": "The contribution is mainly a prompt-engineering workflow, not a clear change to model training, serving, or systems design. It also has no meaningful memory/storage mechanism, and the evaluation appears limited to compliance gains on a narrow class of prompt-optimization tasks.",
      "downgrade_reasons": [
        "No direct memory/storage relevance",
        "Primarily prompt-engineering workflow",
        "Limited systems impact",
        "Evidence appears narrow and model-limited",
        "Crowded area with many prompt optimization variants"
      ],
      "what_would_raise_priority": "Stronger priority would come from broad evaluation across harder real-world workflows, commercial and open models, and evidence that the method materially improves production reliability beyond benchmark compliance scores.",
      "one_sentence_episode_hook": "Can a multi-agent prompt editor turn flaky instruction following into something measurable and reliable, or is this just a more elaborate way to wordsmith prompts?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03144",
      "title": "Self-Verification is All You Need To Pass The Japanese Bar Examination",
      "abstract": "Despite rapid advances in large language models (LLMs), achieving reliable performance on highly professional and structured examinations remains a significant challenge. The Japanese bar examination is a particularly demanding benchmark, requiring not only advanced legal reasoning but also strict adherence to complex answer formats that involve joint evaluation of multiple propositions. While recent studies have reported improvements by decomposing such questions into simpler true--false judgments, these approaches have not been systematically evaluated under the original exam format and scoring scheme, leaving open the question of whether they truly capture exam-level competence. In this paper, we present a self-verification model trained on a newly constructed dataset that faithfully replicates the authentic format and evaluation scale of the exam. Our model is able to exceed the official passing score when evaluated on the actual exam scale, marking the first demonstration, to our knowledge, of an LLM passing the Japanese bar examination without altering its original question structure or scoring rules. We further conduct extensive comparisons with alternative strategies, including multi-agent inference and decomposition-based supervision, and find that these methods fail to achieve comparable performance. Our results highlight the importance of format-faithful supervision and consistency verification, and suggest that carefully designed single-model approaches can outperform more complex systems in high-stakes professional reasoning tasks. Our dataset and codes are publicly available.",
      "authors": [
        "Andrew Shin"
      ],
      "published_at": "2026-01-06T16:13:47",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03144v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.46543243527412415,
      "sim_memory": 0.2781311273574829,
      "sim_negative": 0.1273176372051239,
      "broad_relevance": 0.46543243527412415,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.550918847322464,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.2781311273574829,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.45999999999999996,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.41226755768060686,
      "memory_score": 0.006439338207244849,
      "quality_score": 0.525,
      "bridge_score": 0.006439338207244849,
      "max_axis_score": 0.41226755768060686,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "Claims of an LLM passing a national bar exam are attention-grabbing, and the paper argues that format-faithful supervision plus self-verification can beat more elaborate reasoning setups. It also speaks to a current editorial theme around whether benchmark gains come from better reasoning or just better task framing.",
      "why_not_higher": "This is still a narrow legal-exam result with limited evidence that the method generalizes beyond one structured professional benchmark. It has no meaningful memory/storage contribution and only modest implications for how the broader field builds modern AI systems.",
      "downgrade_reasons": [
        "Narrow task-specific benchmark",
        "Memory/storage link absent",
        "Unclear transfer beyond Japanese bar exam format",
        "Episode risks collapsing into score-chasing rather than broader systems insight"
      ],
      "what_would_raise_priority": "Show the same self-verification approach transferring across multiple high-stakes exams or structured reasoning benchmarks with strong ablations on model size, cost, and robustness.",
      "one_sentence_episode_hook": "Did an LLM really pass the Japanese bar exam\u2014or did careful self-verification and format-faithful training finally expose how much of 'reasoning' is actually exam mechanics?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03423",
      "title": "Training-Free Adaptation of New-Generation LLMs using Legacy Clinical Models",
      "abstract": "Adapting language models to the clinical domain through continued pretraining and fine-tuning requires costly retraining for each new model generation. We propose Cross-Architecture Proxy Tuning (CAPT), a model-ensembling approach that enables training-free adaptation of state-of-the-art general-domain models using existing clinical models. CAPT supports models with disjoint vocabularies, leveraging contrastive decoding to selectively inject clinically relevant signals while preserving the general-domain model's reasoning and fluency. On six clinical classification and text-generation tasks, CAPT with a new-generation general-domain model and an older-generation clinical model consistently outperforms both models individually and state-of-the-art ensembling approaches (average +17.6% over UniTE, +41.4% over proxy tuning across tasks). Through token-level analysis and physician case studies, we demonstrate that CAPT amplifies clinically actionable language, reduces context errors, and increases clinical specificity.",
      "authors": [
        "Sasha Ronaghi",
        "Chloe Stanwyck",
        "Asad Aali",
        "Amir Ronaghi",
        "Miguel Fuentes",
        "Tina Hernandez-Boussard",
        "Emily Alsentzer"
      ],
      "published_at": "2026-01-06T21:23:47",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03423v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.452046275138855,
      "sim_memory": 0.4217870831489563,
      "sim_negative": 0.22705890238285065,
      "broad_relevance": 0.452046275138855,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5662785768508911,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.4217870831489563,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3905556690692902,
      "memory_score": 0.02953612494468691,
      "quality_score": 0.595,
      "bridge_score": 0.02953612494468691,
      "max_axis_score": 0.3905556690692902,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Inference",
        "Application",
        "Bridge"
      ],
      "status": "Deferred this cycle",
      "why_now": "It tackles a timely problem: how to reuse legacy domain models to adapt newer frontier LLMs without retraining, which is attractive for regulated domains like healthcare where domain adaptation lags model releases. The training-free cross-architecture ensembling angle is also operationally relevant because many teams have old specialist models but cannot afford continual re-pretraining.",
      "why_not_higher": "The paper is still mainly a clinical-domain adaptation result rather than a broadly reusable shift in how the field builds or serves LLMs. Its memory/storage relevance is weak, since the core contribution is contrastive proxy ensembling across vocabularies, not cache, bandwidth, offload, locality, or serving systems behavior.",
      "downgrade_reasons": [
        "narrow clinical application focus",
        "memory/storage link is not direct",
        "ensemble adaptation may add inference complexity without systems analysis",
        "unclear transfer beyond clinical domain"
      ],
      "what_would_raise_priority": "Show strong transfer to multiple non-clinical domains and quantify real serving tradeoffs such as latency, memory footprint, and cost versus fine-tuning or standard ensembling.",
      "one_sentence_episode_hook": "Can you bolt an old clinical specialist onto a brand-new general LLM and get domain expertise for free\u2014without retraining either model?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03248",
      "title": "STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning",
      "abstract": "Spatio-temporal reasoning in time series involves the explicit synthesis of temporal dynamics, spatial dependencies, and textual context. This capability is vital for high-stakes decision-making in systems such as traffic networks, power grids, and disease propagation. However, the field remains underdeveloped because most existing works prioritize predictive accuracy over reasoning. To address the gap, we introduce ST-Bench, a benchmark consisting of four core tasks, including etiological reasoning, entity identification, correlation reasoning, and in-context forecasting, developed via a network SDE-based multi-agent data synthesis pipeline. We then propose STReasoner, which empowers LLM to integrate time series, graph structure, and text for explicit reasoning. To promote spatially grounded logic, we introduce S-GRPO, a reinforcement learning algorithm that rewards performance gains specifically attributable to spatial information. Experiments show that STReasoner achieves average accuracy gains between 17% and 135% at only 0.004X the cost of proprietary models and generalizes robustly to real-world data.",
      "authors": [
        "Juntong Ni",
        "Shiyu Wang",
        "Ming Jin",
        "Qi He",
        "Wei Jin"
      ],
      "published_at": "2026-01-06T18:46:12",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03248v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.46927714347839355,
      "sim_memory": 0.3206529915332794,
      "sim_negative": 0.34634795784950256,
      "broad_relevance": 0.46927714347839355,
      "momentum": 0.017490950964499406,
      "teachability": 0.75,
      "novelty_score": 0.5503132343292236,
      "evidence_score": 0.38931471805599455,
      "direct_memory_relevance": 0.3206529915332794,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.38722552609420063,
      "memory_score": 0.034513663626703134,
      "quality_score": 0.6027258872223977,
      "bridge_score": 0.034513663626703134,
      "max_axis_score": 0.38722552609420063,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "It reflects a live trend toward pushing LLMs beyond text-only reasoning into structured spatio-temporal settings, with a new benchmark and an RL-based training recipe. That could interest listeners tracking how benchmarks are being built for more realistic decision-support tasks.",
      "why_not_higher": "The contribution is mostly a task-specific benchmark and method for spatio-temporal reasoning rather than something that clearly changes mainstream model building, serving, or evaluation. Its memory/storage relevance is essentially absent, and the claimed impact appears narrow unless the benchmark and method transfer broadly across domains and model families.",
      "downgrade_reasons": [
        "Memory/storage connection absent",
        "Task-specific spatio-temporal benchmark",
        "Unclear impact on mainstream AI systems practice",
        "Evidence appears benchmark-centric rather than deployment-centric",
        "Generalization claims need stronger real-world validation"
      ],
      "what_would_raise_priority": "Stronger evidence that S-GRPO or the multimodal reasoning setup transfers to widely used real-world forecasting/graph tasks and materially improves standard LLM evaluation practice would raise it.",
      "one_sentence_episode_hook": "Can reinforcement learning teach LLMs to reason over time series plus graphs, or is STReasoner mostly a polished benchmark for a narrow class of problems?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03023",
      "title": "MedDialogRubrics: A Comprehensive Benchmark and Evaluation Framework for Multi-turn Medical Consultations in Large Language Models",
      "abstract": "Medical conversational AI (AI) plays a pivotal role in the development of safer and more effective medical dialogue systems. However, existing benchmarks and evaluation frameworks for assessing the information-gathering and diagnostic reasoning abilities of medical large language models (LLMs) have not been rigorously evaluated. To address these gaps, we present MedDialogRubrics, a novel benchmark comprising 5,200 synthetically constructed patient cases and over 60,000 fine-grained evaluation rubrics generated by LLMs and subsequently refined by clinical experts, specifically designed to assess the multi-turn diagnostic capabilities of LLM. Our framework employs a multi-agent system to synthesize realistic patient records and chief complaints from underlying disease knowledge without accessing real-world electronic health records, thereby mitigating privacy and data-governance concerns. We design a robust Patient Agent that is limited to a set of atomic medical facts and augmented with a dynamic guidance mechanism that continuously detects and corrects hallucinations throughout the dialogue, ensuring internal coherence and clinical plausibility of the simulated cases. Furthermore, we propose a structured LLM-based and expert-annotated rubric-generation pipeline that retrieves Evidence-Based Medicine (EBM) guidelines and utilizes the reject sampling to derive a prioritized set of rubric items (\"must-ask\" items) for each case. We perform a comprehensive evaluation of state-of-the-art models and demonstrate that, across multiple assessment dimensions, current models face substantial challenges. Our results indicate that improving medical dialogue will require advances in dialogue management architectures, not just incremental tuning of the base-model.",
      "authors": [
        "Lecheng Gong",
        "Weimin Fang",
        "Ting Yang",
        "Dongjie Tao",
        "Chunxiao Guo",
        "Peng Wei",
        "Bo Xie",
        "Jinqun Guan",
        "Zixiao Chen",
        "Fang Shi"
      ],
      "published_at": "2026-01-06T13:56:33",
      "categories": [
        "cs.CL",
        "cs.HC"
      ],
      "url": "http://arxiv.org/abs/2601.03023v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4281139373779297,
      "sim_memory": 0.2651517987251282,
      "sim_negative": 0.27549493312835693,
      "broad_relevance": 0.4281139373779297,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.6082450151443481,
      "evidence_score": 0.5293147180559945,
      "direct_memory_relevance": 0.2651517987251282,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.38356633138633023,
      "memory_score": 0.017863305784257794,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.017863305784257794,
      "max_axis_score": 0.38356633138633023,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "Medical LLM evaluation is timely, and this paper offers a fairly substantial benchmark for multi-turn diagnostic dialogue with expert-refined rubrics. It also highlights that current models still struggle with clinically grounded information gathering, which is a useful reality check for conversational AI.",
      "why_not_higher": "The contribution is mainly a domain-specific benchmark and evaluation framework rather than a method that changes mainstream LLM training, inference, or systems practice. Its memory/storage relevance is effectively absent, and the transfer to broader AI is limited.",
      "downgrade_reasons": [
        "narrow medical dialogue focus",
        "benchmark more than method",
        "no direct memory/storage contribution",
        "limited impact on how modern AI systems are built or served",
        "synthetic workload may not fully capture real clinical deployment"
      ],
      "what_would_raise_priority": "Show that the rubric framework or dialogue architecture insights transfer beyond medicine to general multi-turn agent evaluation or to production conversational system design.",
      "one_sentence_episode_hook": "What if today\u2019s best LLMs can sound medically fluent but still fail the basic question-asking discipline a real diagnosis requires?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03436",
      "title": "MARVEL: A Multi Agent-based Research Validator and Enabler using Large Language Models",
      "abstract": "We present MARVEL (https://ligogpt.mit.edu/marvel), a locally deployable, open-source framework for domain-aware question answering and assisted scientific research. It is designed to address the increasing demands of a digital assistant for scientific groups that can read highly technical data, cite precisely, and operate within authenticated networks. MARVEL combines a fast path for straightforward queries with a more deliberate DeepSearch mode that integrates retrieval-augmented generation and Monte Carlo Tree Search. It explores complementary subqueries, allocates more compute to promising branches, and maintains a global evidence ledger that preserves sources during drafting. We applied this framework in the context of gravitational-wave research related to the Laser Interferometer Gravitational-wave Observatory. Answers are grounded in a curated semantic index of research literature, doctoral theses, LIGO documents, and long-running detector electronic logbooks, with targeted web searches when appropriate. Because direct benchmarking against commercial LLMs cannot be performed on private data, we evaluated MARVEL on two publicly available surrogate datasets that capture comparable semantic and technical characteristics. On these benchmarks, MARVEL matches a GPT-4o mini baseline on literature-centric queries and substantially outperforms it on detector-operations content, where domain retrieval and guided reasoning are decisive. By making the complete framework and evaluation datasets openly available, we aim to provide a reproducible foundation for developing domain-specific scientific assistants.",
      "authors": [
        "Nikhil Mukund",
        "Yifang Luo",
        "Fan Zhang",
        "Lisa Barsotti",
        "Erik Katsavounidis"
      ],
      "published_at": "2026-01-06T21:47:22",
      "categories": [
        "astro-ph.IM",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03436v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4092511534690857,
      "sim_memory": 0.35749727487564087,
      "sim_negative": 0.23873060941696167,
      "broad_relevance": 0.4092511534690857,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6143798530101776,
      "evidence_score": 0.38,
      "direct_memory_relevance": 0.35749727487564087,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3749323239922524,
      "memory_score": 0.023249182462692253,
      "quality_score": 0.575,
      "bridge_score": 0.023249182462692253,
      "max_axis_score": 0.3749323239922524,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "Scientific groups are actively trying to deploy local, domain-aware assistants inside private networks, and this paper offers an open-source blueprint with retrieval, citation tracking, and a deeper search mode. It is timely as labs want alternatives to closed commercial assistants for sensitive technical workflows.",
      "why_not_higher": "The contribution is mostly an application-specific orchestration stack for gravitational-wave research rather than a broadly transferable advance in modern AI system design. Its memory/storage relevance is incidental, and the evaluation relies on surrogate benchmarks rather than direct head-to-head testing on the private target workload.",
      "downgrade_reasons": [
        "narrow domain focus",
        "limited transfer beyond scientific QA assistants",
        "not a memory/storage paper",
        "evaluation on surrogate datasets",
        "framework integration more than core model advance"
      ],
      "what_would_raise_priority": "Stronger evidence that the DeepSearch and evidence-ledger design generalize across multiple scientific domains and materially improve real-world evaluation over standard RAG baselines would raise its priority.",
      "one_sentence_episode_hook": "Can a locally deployable multi-agent assistant actually help scientists navigate private technical archives better than a generic LLM, or is MARVEL mainly a well-engineered domain RAG stack?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03450",
      "title": "Soft Contextualized Encoder For User Defined Text Classification",
      "abstract": "User-Defined Text Classification (UDTC) considers the challenge of classifying input text to user-specified, previously unseen classes, a setting that arises frequently in real-world applications such as enterprise analytics, content moderation, and domain-specific information retrieval. We propose a soft-contextualized encoder architecture for UDTC which contextualizes each candidate label with the label set and a static soft prompt representation of the input query. Training on diverse, multi-source datasets enables the model to generalize effectively to zero-shot classification over entirely unseen topic sets drawn from arbitrary domains. We evaluate the proposed architecture both on held-out in-distribution test data and on multiple unseen UDTC benchmarks. Across datasets, the model achieves state-of-the-art performance, consistently outperforming or matching the baselines.",
      "authors": [
        "Charu Maheshwari",
        "Vyas Raina"
      ],
      "published_at": "2026-01-06T22:34:50",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03450v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.37375035881996155,
      "sim_memory": 0.28140151500701904,
      "sim_negative": 0.2644392251968384,
      "broad_relevance": 0.37375035881996155,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5909383594989777,
      "evidence_score": 0.37,
      "direct_memory_relevance": 0.28140151500701904,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.37076586157083513,
      "memory_score": 0.0004204545021057049,
      "quality_score": 0.595,
      "bridge_score": 0.0004204545021057049,
      "max_axis_score": 0.37076586157083513,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "Zero-shot user-defined classification is practically useful for enterprise and moderation-style workflows, and the paper claims strong performance on unseen label sets across multiple datasets. It may interest listeners tracking pragmatic alternatives to full supervised retraining for changing taxonomies.",
      "why_not_higher": "This is still a fairly narrow task paper centered on UDTC benchmarks rather than a broader shift in how modern AI systems are built or served. The memory/storage connection is effectively absent, and the methodological novelty appears incremental rather than field-shaping.",
      "downgrade_reasons": [
        "narrow task focus",
        "no direct memory/storage mechanism",
        "limited systems relevance",
        "primarily benchmark-driven contribution",
        "unclear broad adoption impact in 30-90 days"
      ],
      "what_would_raise_priority": "Show that the approach generalizes into a widely adopted classification/retrieval serving pattern or yields clear efficiency, latency, or deployment advantages at production scale.",
      "one_sentence_episode_hook": "Can a softly contextualized encoder make zero-shot, user-defined text classification practical enough to replace constant retraining when label sets keep changing?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03194",
      "title": "X-MuTeST: A Multilingual Benchmark for Explainable Hate Speech Detection and A Novel LLM-consulted Explanation Framework",
      "abstract": "Hate speech detection on social media faces challenges in both accuracy and explainability, especially for underexplored Indic languages. We propose a novel explainability-guided training framework, X-MuTeST (eXplainable Multilingual haTe Speech deTection), for hate speech detection that combines high-level semantic reasoning from large language models (LLMs) with traditional attention-enhancing techniques. We extend this research to Hindi and Telugu alongside English by providing benchmark human-annotated rationales for each word to justify the assigned class label. The X-MuTeST explainability method computes the difference between the prediction probabilities of the original text and those of unigrams, bigrams, and trigrams. Final explanations are computed as the union between LLM explanations and X-MuTeST explanations. We show that leveraging human rationales during training enhances both classification performance and explainability. Moreover, combining human rationales with our explainability method to refine the model attention yields further improvements. We evaluate explainability using Plausibility metrics such as Token-F1 and IOU-F1 and Faithfulness metrics such as Comprehensiveness and Sufficiency. By focusing on under-resourced languages, our work advances hate speech detection across diverse linguistic contexts. Our dataset includes token-level rationale annotations for 6,004 Hindi, 4,492 Telugu, and 6,334 English samples. Data and code are available on https://github.com/ziarehman30/X-MuTeST",
      "authors": [
        "Mohammad Zia Ur Rehman",
        "Sai Kartheek Reddy Kasu",
        "Shashivardhan Reddy Koppula",
        "Sai Rithwik Reddy Chirra",
        "Shwetank Shekhar Singh",
        "Nagendra Kumar"
      ],
      "published_at": "2026-01-06T17:16:45",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03194v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4252302050590515,
      "sim_memory": 0.3682069778442383,
      "sim_negative": 0.21635477244853973,
      "broad_relevance": 0.4252302050590515,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6189889907836914,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.3682069778442383,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.37041741013526913,
      "memory_score": 0.0,
      "quality_score": 0.595,
      "bridge_score": 0.0,
      "max_axis_score": 0.37041741013526913,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "It targets a timely problem\u2014multilingual, explainable hate-speech detection in under-resourced Indic languages\u2014and contributes token-level rationale data plus an LLM-assisted explanation pipeline. The explainability angle and released benchmark could interest practitioners working on safety and moderation datasets.",
      "why_not_higher": "This is mainly a narrow task-and-benchmark paper rather than something that changes how the broader field builds or serves modern AI systems. Its memory/storage relevance is effectively absent, and the proposed method does not obviously generalize into a widely adopted systems or foundation-model practice.",
      "downgrade_reasons": [
        "Narrow application focus",
        "Benchmark-centric contribution",
        "Weak transfer beyond hate-speech detection",
        "No direct memory/storage mechanism",
        "Limited systems relevance for modern AI stacks"
      ],
      "what_would_raise_priority": "Evidence that the explanation framework transfers across multiple safety/NLP tasks and materially improves mainstream LLM or classifier evaluation workflows would raise its priority.",
      "one_sentence_episode_hook": "Can LLM-generated rationales and human token annotations make hate-speech detection in Hindi and Telugu both more accurate and more explainable\u2014or is this still just a niche benchmark win?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02983",
      "title": "Interpretable All-Type Audio Deepfake Detection with Audio LLMs via Frequency-Time Reinforcement Learning",
      "abstract": "Recent advances in audio large language models (ALLMs) have made high-quality synthetic audio widely accessible, increasing the risk of malicious audio deepfakes across speech, environmental sounds, singing voice, and music. Real-world audio deepfake detection (ADD) therefore requires all-type detectors that generalize across heterogeneous audio and provide interpretable decisions. Given the strong multi-task generalization ability of ALLMs, we first investigate their performance on all-type ADD under both supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT). However, SFT using only binary real/fake labels tends to reduce the model to a black-box classifier, sacrificing interpretability. Meanwhile, vanilla RFT under sparse supervision is prone to reward hacking and can produce hallucinated, ungrounded rationales. To address this, we propose an automatic annotation and polishing pipeline that constructs Frequency-Time structured chain-of-thought (CoT) rationales, producing ~340K cold-start demonstrations. Building on CoT data, we propose Frequency Time-Group Relative Policy Optimization (FT-GRPO), a two-stage training paradigm that cold-starts ALLMs with SFT and then applies GRPO under rule-based frequency-time constraints. Experiments demonstrate that FT-GRPO achieves state-of-the-art performance on all-type ADD while producing interpretable, FT-grounded rationales. The data and code are available online.",
      "authors": [
        "Yuankun Xie",
        "Xiaoxuan Guo",
        "Jiayi Zhou",
        "Tao Wang",
        "Jian Liu",
        "Ruibo Fu",
        "Xiaopeng Wang",
        "Haonan Cheng",
        "Long Ye"
      ],
      "published_at": "2026-01-06T12:50:02",
      "categories": [
        "cs.SD",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.02983v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 4,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "audio",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.44100192189216614,
      "sim_memory": 0.3582763671875,
      "sim_negative": 0.2335236370563507,
      "broad_relevance": 0.44100192189216614,
      "momentum": 0.05256382402428469,
      "teachability": 0.7,
      "novelty_score": 0.6338478326797485,
      "evidence_score": 0.42094379124341,
      "direct_memory_relevance": 0.3582763671875,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.5,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.3,
      "public_interest_score": 0.36703208496098055,
      "memory_score": 0.017796165105459172,
      "quality_score": 0.5393775164973641,
      "bridge_score": 0.017796165105459172,
      "max_axis_score": 0.36703208496098055,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "Audio deepfakes are becoming more practical across speech, music, and environmental audio, so an interpretable detector spanning multiple audio types is timely. The paper also taps into a current trend of using LLM-style rationale generation plus RL-style fine-tuning for trust and robustness.",
      "why_not_higher": "This is still primarily a narrow audio deepfake detection paper rather than a result that broadly changes how modern AI systems are built or served. Its memory/storage relevance is essentially absent, and the interpretability claims rely on a task-specific rationale pipeline rather than a clearly general systems contribution.",
      "downgrade_reasons": [
        "narrow application focus",
        "memory/storage connection absent",
        "limited systems relevance",
        "interpretability method appears domain-specific",
        "unclear transfer beyond audio deepfake detection"
      ],
      "what_would_raise_priority": "Stronger evidence that the frequency-time rationale and FT-GRPO framework transfer to multiple multimodal detection or attribution tasks beyond audio deepfakes would raise priority.",
      "one_sentence_episode_hook": "Can audio LLMs explain exactly where a deepfake hides in time and frequency, or are we just training better-sounding black boxes?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03162",
      "title": "On the Convergence Behavior of Preconditioned Gradient Descent Toward the Rich Learning Regime",
      "abstract": "Spectral bias, the tendency of neural networks to learn low frequencies first, can be both a blessing and a curse. While it enhances the generalization capabilities by suppressing high-frequency noise, it can be a limitation in scientific tasks that require capturing fine-scale structures. The delayed generalization phenomenon known as grokking is another barrier to rapid training of neural networks. Grokking has been hypothesized to arise as learning transitions from the NTK to the feature-rich regime. This paper explores the impact of preconditioned gradient descent (PGD), such as Gauss-Newton, on spectral bias and grokking phenomena. We demonstrate through theoretical and empirical results how PGD can mitigate issues associated with spectral bias. Additionally, building on the rich learning regime grokking hypothesis, we study how PGD can be used to reduce delays associated with grokking. Our conjecture is that PGD, without the impediment of spectral bias, enables uniform exploration of the parameter space in the NTK regime. Our experimental results confirm this prediction, providing strong evidence that grokking represents a transitional behavior between the lazy regime characterized by the NTK and the rich regime. These findings deepen our understanding of the interplay between optimization dynamics, spectral bias, and the phases of neural network learning.",
      "authors": [
        "Shuai Jiang",
        "Alexey Voronin",
        "Eric Cyr",
        "Ben Southworth"
      ],
      "published_at": "2026-01-06T16:35:07",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03162v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "theory",
      "narrow_domain_flag": false,
      "sim_public": 0.4762493669986725,
      "sim_memory": 0.5134687423706055,
      "sim_negative": 0.32585448026657104,
      "broad_relevance": 0.4762493669986725,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.4936806559562683,
      "evidence_score": 0.24931471805599453,
      "direct_memory_relevance": 0.5134687423706055,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7499999999999999,
      "clarity": 0.6,
      "reproducibility": 0.3,
      "public_interest_score": 0.365822306394341,
      "memory_score": 0.026358388877900957,
      "quality_score": 0.5127258872223979,
      "bridge_score": 0.026358388877900957,
      "max_axis_score": 0.365822306394341,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "It tackles two conceptually interesting training phenomena, spectral bias and grokking, and argues that preconditioned gradient methods can change the transition from lazy to rich learning. That could matter to researchers thinking about optimization dynamics and representation learning.",
      "why_not_higher": "The paper appears primarily theoretical and small-scale empirical rather than tied to realistic frontier model training or deployment. Its connection to memory/storage is effectively absent, and the practical path from these results to how teams build modern systems is still unclear.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Optimizer/theory paper without clear large-model systems relevance",
        "Unclear evidence on realistic workloads or hardware",
        "Limited immediate impact on current model building or serving"
      ],
      "what_would_raise_priority": "Show that the proposed preconditioning reliably improves training speed or generalization on modern large-scale transformer workloads with realistic compute and memory constraints.",
      "one_sentence_episode_hook": "Can better preconditioning break spectral bias and explain grokking as the moment a network escapes lazy NTK learning into a richer regime?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03034",
      "title": "NorwAI's Large Language Models: Technical Report",
      "abstract": "Norwegian, spoken by approximately five million people, remains underrepresented in many of the most significant breakthroughs in Natural Language Processing (NLP). To address this gap, the NorLLM team at NorwAI has developed a family of models specifically tailored to Norwegian and other Scandinavian languages, building on diverse Transformer-based architectures such as GPT, Mistral, Llama2, Mixtral and Magistral. These models are either pretrained from scratch or continually pretrained on 25B - 88.45B tokens, using a Norwegian-extended tokenizer and advanced post-training strategies to optimize performance, enhance robustness, and improve adaptability across various real-world tasks. Notably, instruction-tuned variants (e.g., Mistral-7B-Instruct and Mixtral-8x7B-Instruct) showcase strong assistant-style capabilities, underscoring their potential for practical deployment in interactive and domain-specific applications. The NorwAI large language models are openly available to Nordic organizations, companies and students for both research and experimental use. This report provides detailed documentation of the model architectures, training data, tokenizer design, fine-tuning strategies, deployment, and evaluations.",
      "authors": [
        "Jon Atle Gulla",
        "Peng Liu",
        "Lemei Zhang"
      ],
      "published_at": "2026-01-06T14:06:55",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03034v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.5713204741477966,
      "sim_memory": 0.5394403338432312,
      "sim_negative": 0.2754025459289551,
      "broad_relevance": 0.5713204741477966,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.529296875,
      "evidence_score": 0.27,
      "direct_memory_relevance": 0.5394403338432312,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.54,
      "clarity": 0.7,
      "reproducibility": 0.3,
      "public_interest_score": 0.36579067349433897,
      "memory_score": 0.045832100152969346,
      "quality_score": 0.455,
      "bridge_score": 0.045832100152969346,
      "max_axis_score": 0.36579067349433897,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Public AI",
        "Training",
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "Open technical reports on region-specific LLM families are useful for documenting how smaller-language ecosystems build viable instruction-tuned models and tokenizer extensions. It also reflects the ongoing push to localize foundation models for underrepresented languages.",
      "why_not_higher": "This is mainly a model family report for Norwegian and Scandinavian use rather than a broadly transferable advance in AI methods or systems. The memory/storage angle is weak, and the main contribution is coverage and engineering integration rather than a new technique likely to shift the field quickly.",
      "downgrade_reasons": [
        "regional model report more than general method paper",
        "memory/storage connection absent in a first-class sense",
        "limited evidence of field-changing systems or evaluation innovation",
        "episode risks becoming a survey of benchmarks and model variants"
      ],
      "what_would_raise_priority": "Clear, reusable lessons on tokenizer design, continual pretraining, and post-training that materially improve small-language LLM development across other low-resource settings would raise its priority.",
      "one_sentence_episode_hook": "What does it actually take to build a serious open LLM ecosystem for a five-million-speaker language without just copying the big-model playbook?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03103",
      "title": "Who Laughs with Whom? Disentangling Influential Factors in Humor Preferences across User Clusters and LLMs",
      "abstract": "Humor preferences vary widely across individuals and cultures, complicating the evaluation of humor using large language models (LLMs). In this study, we model heterogeneity in humor preferences in Oogiri, a Japanese creative response game, by clustering users with voting logs and estimating cluster-specific weights over interpretable preference factors using Bradley-Terry-Luce models. We elicit preference judgments from LLMs by prompting them to select the funnier response and found that user clusters exhibit distinct preference patterns and that the LLM results can resemble those of particular clusters. Finally, we demonstrate that, by persona prompting, LLM preferences can be directed toward a specific cluster. The scripts for data collection and analysis will be released to support reproducibility.",
      "authors": [
        "Soichiro Murakami",
        "Hidetaka Kamigaito",
        "Hiroya Takamura",
        "Manabu Okumura"
      ],
      "published_at": "2026-01-06T15:33:45",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03103v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.377243310213089,
      "sim_memory": 0.29500865936279297,
      "sim_negative": 0.31516438722610474,
      "broad_relevance": 0.377243310213089,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6766054630279541,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.29500865936279297,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.52,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.36466381251811986,
      "memory_score": 0.004502597808837883,
      "quality_score": 0.525,
      "bridge_score": 0.004502597808837883,
      "max_axis_score": 0.36466381251811986,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "It touches a timely question for LLM evaluation: whose preferences a model is actually matching, and whether persona prompting can steer judgments toward specific audience clusters. That makes it relevant to ongoing debates about subjective evaluation and personalization.",
      "why_not_higher": "The paper is centered on humor preference modeling in a specific Japanese Oogiri setting, which is interesting but narrow for a broad AI audience. It also has no direct memory/storage contribution and only limited impact on how mainstream systems are built or served.",
      "downgrade_reasons": [
        "narrow task domain",
        "subjective evaluation focus",
        "no direct memory/storage mechanism",
        "limited systems relevance",
        "unclear generalization beyond humor/Oogiri"
      ],
      "what_would_raise_priority": "Show that the cluster-based preference modeling and persona steering reliably transfer to broader LLM evaluation domains such as helpfulness, style, safety, or recommendation at production scale.",
      "one_sentence_episode_hook": "When an LLM says a joke is funny, is it reflecting universal taste\u2014or just imitating one hidden audience cluster?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.04252",
      "title": "Sphinx: Benchmarking and Modeling for LLM-Driven Pull Request Review",
      "abstract": "Pull request (PR) review is essential for ensuring software quality, yet automating this task remains challenging due to noisy supervision, limited contextual understanding, and inadequate evaluation metrics. We present Sphinx, a unified framework for LLM-based PR review that addresses these limitations through three key components: (1) a structured data generation pipeline that produces context-rich, semantically grounded review comments by comparing pseudo-modified and merged code; (2) a checklist-based evaluation benchmark that assesses review quality based on structured coverage of actionable verification points, moving beyond surface-level metrics like BLEU; and (3) Checklist Reward Policy Optimization (CRPO), a novel training paradigm that uses rule-based, interpretable rewards to align model behavior with real-world review practices. Extensive experiments show that models trained with Sphinx achieve state-of-the-art performance on review completeness and precision, outperforming both proprietary and open-source baselines by up to 40\\% in checklist coverage. Together, Sphinx enables the development of PR review models that are not only fluent but also context-aware, technically precise, and practically deployable in real-world development workflows. The data will be released after review.",
      "authors": [
        "Daoan Zhang",
        "Shuo Zhang",
        "Zijian Jin",
        "Jiebo Luo",
        "Shengyu Fu",
        "Elsie Nallipogu"
      ],
      "published_at": "2026-01-06T18:49:56",
      "categories": [
        "cs.SE",
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.04252v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.40398773550987244,
      "sim_memory": 0.2906057834625244,
      "sim_negative": 0.2668013870716095,
      "broad_relevance": 0.40398773550987244,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6171218156814575,
      "evidence_score": 0.38,
      "direct_memory_relevance": 0.2906057834625244,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3637645930051804,
      "memory_score": 0.0,
      "quality_score": 0.575,
      "bridge_score": 0.0,
      "max_axis_score": 0.3637645930051804,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "LLM-based software engineering is a live topic, and PR review is a credible workflow where better evaluation and reward design could matter. The checklist benchmark angle is timely because many code-assistant papers still rely on weak surface metrics.",
      "why_not_higher": "This is still a fairly narrow task paper centered on pull request review rather than a broadly reusable advance in how modern AI systems are built or served. It also has no real memory/storage story beyond incidental LLM usage.",
      "downgrade_reasons": [
        "narrow software-engineering task",
        "memory/storage connection absent",
        "unclear generalization beyond PR review",
        "benchmark and training recipe more than platform-level shift",
        "claims appear strong but realism of deployment setting is not fully clear from abstract"
      ],
      "what_would_raise_priority": "Show that the checklist/reward framework transfers cleanly to multiple code-assistant tasks and improves real developer outcomes in live review workflows.",
      "one_sentence_episode_hook": "Can we teach LLMs to do useful pull-request review by rewarding checklist coverage instead of pretty-sounding comments?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03329",
      "title": "Attention mechanisms in neural networks",
      "abstract": "Attention mechanisms represent a fundamental paradigm shift in neural network architectures, enabling models to selectively focus on relevant portions of input sequences through learned weighting functions. This monograph provides a comprehensive and rigorous mathematical treatment of attention mechanisms, encompassing their theoretical foundations, computational properties, and practical implementations in contemporary deep learning systems. Applications in natural language processing, computer vision, and multimodal learning demonstrate the versatility of attention mechanisms. We examine language modeling with autoregressive transformers, bidirectional encoders for representation learning, sequence-to-sequence translation, Vision Transformers for image classification, and cross-modal attention for vision-language tasks. Empirical analysis reveals training characteristics, scaling laws that relate performance to model size and computation, attention pattern visualizations, and performance benchmarks across standard datasets. We discuss the interpretability of learned attention patterns and their relationship to linguistic and visual structures. The monograph concludes with a critical examination of current limitations, including computational scalability, data efficiency, systematic generalization, and interpretability challenges.",
      "authors": [
        "Hasi Hays"
      ],
      "published_at": "2026-01-06T17:12:10",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03329v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 5,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.587158203125,
      "sim_memory": 0.4430606961250305,
      "sim_negative": 0.4127056896686554,
      "broad_relevance": 0.587158203125,
      "momentum": 0.06108604879161034,
      "teachability": 0.75,
      "novelty_score": 0.3839256167411804,
      "evidence_score": 0.4791759469228055,
      "direct_memory_relevance": 0.4430606961250305,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.36282990524542,
      "memory_score": 0.0574193224682458,
      "quality_score": 0.6466703787691221,
      "bridge_score": 0.0574193224682458,
      "max_axis_score": 0.36282990524542,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory"
      ],
      "status": "Deferred this cycle",
      "why_now": "Attention remains foundational to modern AI, so a rigorous monograph can be educational for listeners who want a structured overview of the field. It may serve as a refresher amid continued transformer-centric research and deployment.",
      "why_not_higher": "This appears to be a broad survey/monograph rather than a concrete new result that changes current practice in building, serving, or evaluating AI systems. Its memory angle is mostly indirect through generic scalability discussion, not a first-class systems contribution on KV cache, bandwidth, offload, or storage behavior.",
      "downgrade_reasons": [
        "survey-like monograph rather than a clear new contribution",
        "memory/storage connection is indirect",
        "unclear realistic systems or hardware evidence",
        "does not obviously change current practice within 30-90 days",
        "high overlap with well-known transformer background material"
      ],
      "what_would_raise_priority": "A stronger showing of new empirical or systems findings on modern large-scale transformer workloads, especially with concrete memory/cache implications, would raise priority.",
      "one_sentence_episode_hook": "If attention is the engine of modern AI, does this monograph actually tell us something new about how it works in practice\u2014or just organize what we already know?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03120",
      "title": "A framework for assuring the accuracy and fidelity of an AI-enabled Digital Twin of en route UK airspace",
      "abstract": "Digital Twins combine simulation, operational data and Artificial Intelligence (AI), and have the potential to bring significant benefits across the aviation industry. Project Bluebird, an industry-academic collaboration, has developed a probabilistic Digital Twin of en route UK airspace as an environment for training and testing AI Air Traffic Control (ATC) agents. There is a developing regulatory landscape for this kind of novel technology. Regulatory requirements are expected to be application specific, and may need to be tailored to each specific use case.   We draw on emerging guidance for both Digital Twin development and the use of Artificial Intelligence/Machine Learning (AI/ML) in Air Traffic Management (ATM) to present an assurance framework. This framework defines actionable goals and the evidence required to demonstrate that a Digital Twin accurately represents its physical counterpart and also provides sufficient functionality across target use cases. It provides a structured approach for researchers to assess, understand and document the strengths and limitations of the Digital Twin, whilst also identifying areas where fidelity could be improved. Furthermore, it serves as a foundation for engagement with stakeholders and regulators, supporting discussions around the regulatory needs for future applications, and contributing to the emerging guidance through a concrete, working example of a Digital Twin.   The framework leverages a methodology known as Trustworthy and Ethical Assurance (TEA) to develop an assurance case. An assurance case is a nested set of structured arguments that provides justified evidence for how a top-level goal has been realised. In this paper we provide an overview of each structured argument and a number of deep dives which elaborate in more detail upon particular arguments, including the required evidence, assumptions and justifications.",
      "authors": [
        "Adam Keane",
        "Nick Pepper",
        "Chris Burr",
        "Amy Hodgkin",
        "Dewi Gould",
        "John Korna",
        "Marc Thomas"
      ],
      "published_at": "2026-01-06T15:49:12",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03120v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 4,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "pde",
      "paper_type": "survey",
      "narrow_domain_flag": true,
      "sim_public": 0.38563746213912964,
      "sim_memory": 0.21636110544204712,
      "sim_negative": 0.2517750859260559,
      "broad_relevance": 0.38563746213912964,
      "momentum": 0.05256382402428469,
      "teachability": 0.75,
      "novelty_score": 0.6248884201049805,
      "evidence_score": 0.40094379124341,
      "direct_memory_relevance": 0.21636110544204712,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.42,
      "clarity": 0.7,
      "reproducibility": 0.3,
      "public_interest_score": 0.3590788351488544,
      "memory_score": 0.008221586581823315,
      "quality_score": 0.494377516497364,
      "bridge_score": 0.008221586581823315,
      "max_axis_score": 0.3590788351488544,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "Assurance and regulatory frameworks for AI-enabled digital twins are becoming more important as safety-critical industries explore AI deployment. This paper offers a concrete case study in air traffic management, a domain where verification and fidelity matter a great deal.",
      "why_not_higher": "It is primarily a domain-specific assurance framework rather than a paper that changes mainstream AI model building, serving, or evaluation. The connection to memory/storage systems is essentially absent, and the transfer beyond aviation digital twins is limited.",
      "downgrade_reasons": [
        "narrow domain focus",
        "framework/assurance paper rather than technical AI advance",
        "no direct memory or storage contribution",
        "limited evidence on broadly relevant modern AI workloads",
        "weak transferability to mainstream transformer practice"
      ],
      "what_would_raise_priority": "Show a broadly reusable assurance methodology validated across multiple AI-enabled simulation systems with concrete implications for evaluating modern foundation-model-based agents.",
      "one_sentence_episode_hook": "What would it take to prove that an AI-powered digital twin of UK airspace is trustworthy enough to train the next generation of air traffic control agents?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03019",
      "title": "DNACHUNKER: Learnable Tokenization for DNA Language Models",
      "abstract": "DNA language models are increasingly used to represent genomic sequence, yet their effectiveness depends critically on how raw nucleotides are converted into model inputs. Unlike natural language, DNA offers no canonical boundaries, making fixed tokenizations a brittle design choice under shifts, indels, and local repeats. We introduce \\modelname{}, a masked DNA language model that incorporates a learnable adaptive segmentation module to produce context-dependent, variable-length units. Building on a dynamic segmentation procedure, \\modelname{} learns to allocate finer granularity to functionally enriched regions while compressing repetitive or redundant sequence. We pre-train \\modelname{} on the human reference genome (HG38) and evaluate it on the Nucleotide Transformer and Genomic Benchmarks, where it consistently improves over strong fixed-tokenization baselines. Further analyses and ablations indicate that the learned segmentation is structured rather than incidental: the model preferentially uses shorter units around promoters and exons, and longer units in repetitive regions, yielding representations that are both mutation-resilient and biologically-informed.",
      "authors": [
        "Taewon Kim",
        "Jihwan Shin",
        "Hyomin Kim",
        "Youngmok Jung",
        "Jonghoon Lee",
        "Won-Chul Lee",
        "Sungsoo Ahn",
        "Insu Han"
      ],
      "published_at": "2026-01-06T13:46:42",
      "categories": [
        "q-bio.GN",
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03019v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.46479612588882446,
      "sim_memory": 0.42765262722969055,
      "sim_negative": 0.2731417119503021,
      "broad_relevance": 0.46479612588882446,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5949121713638306,
      "evidence_score": 0.42000000000000004,
      "direct_memory_relevance": 0.42765262722969055,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.48,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.35867566347122193,
      "memory_score": 0.02429578816890715,
      "quality_score": 0.525,
      "bridge_score": 0.02429578816890715,
      "max_axis_score": 0.35867566347122193,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "It targets a real bottleneck in genomic language modeling: fixed tokenization is a poor fit for DNA, and learnable segmentation is a timely idea as domain-specific foundation models mature. The paper also offers interpretable evidence that the learned units align with biological structure rather than acting as a purely opaque preprocessing trick.",
      "why_not_higher": "The contribution is mostly a domain-specific modeling improvement for DNA rather than a change to mainstream AI practice. Its memory/storage relevance is weak: adaptive tokenization may affect sequence length indirectly, but the paper is not about KV cache, offload, paging, bandwidth, or serving systems.",
      "downgrade_reasons": [
        "Narrow domain focus in genomics",
        "Memory/storage connection is indirect at best",
        "Unclear impact on mainstream LLM build/serve workflows",
        "Primarily a benchmark/modeling paper rather than a systems paper"
      ],
      "what_would_raise_priority": "Show that the adaptive tokenization transfers beyond DNA or produces concrete training/inference efficiency gains on long-context hardware-constrained workloads.",
      "one_sentence_episode_hook": "What if tokenization stopped being a fixed preprocessing choice and became a learnable part of the model\u2014especially in domains like DNA where the 'words' were never real to begin with?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03217",
      "title": "MalruleLib: Large-Scale Executable Misconception Reasoning with Step Traces for Modeling Student Thinking in Mathematics",
      "abstract": "Student mistakes in mathematics are often systematic: a learner applies a coherent but wrong procedure and repeats it across contexts. We introduce MalruleLib, a learning-science-grounded framework that translates documented misconceptions into executable procedures, drawing on 67 learning-science and mathematics education sources, and generates step-by-step traces of malrule-consistent student work. We formalize a core student-modeling problem as Malrule Reasoning Accuracy (MRA): infer a misconception from one worked mistake and predict the student's next answer under cross-template rephrasing. Across nine language models (4B-120B), accuracy drops from 66% on direct problem solving to 40% on cross-template misconception prediction. MalruleLib encodes 101 malrules over 498 parameterized problem templates and produces paired dual-path traces for both correct reasoning and malrule-consistent student reasoning. Because malrules are executable and templates are parameterizable, MalruleLib can generate over one million instances, enabling scalable supervision and controlled evaluation. Using MalruleLib, we observe cross-template degradations of 10-21%, while providing student step traces improves prediction by 3-15%. We release MalruleLib as infrastructure for educational AI that models student procedures across contexts, enabling diagnosis and feedback that targets the underlying misconception.",
      "authors": [
        "Xinghe Chen",
        "Naiming Liu",
        "Shashank Sonkar"
      ],
      "published_at": "2026-01-06T17:59:37",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03217v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.3484623432159424,
      "sim_memory": 0.32161015272140503,
      "sim_negative": 0.1553615927696228,
      "broad_relevance": 0.3484623432159424,
      "momentum": 0.017490950964499406,
      "teachability": 0.75,
      "novelty_score": 0.5915031433105469,
      "evidence_score": 0.4993147180559946,
      "direct_memory_relevance": 0.32161015272140503,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3571595723626639,
      "memory_score": 0.007800811983140865,
      "quality_score": 0.6027258872223977,
      "bridge_score": 0.007800811983140865,
      "max_axis_score": 0.3571595723626639,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "It surfaces a real weakness in current LLMs: they can solve a math problem directly much better than they can infer and generalize a student's underlying misconception across rephrasings. The executable malrule framework and large synthetic trace generation make it a useful educational-AI evaluation resource.",
      "why_not_higher": "This is primarily a student-modeling benchmark/dataset paper rather than something that changes mainstream AI system design, training, serving, or evaluation broadly. Its memory/storage relevance is essentially absent, and the transfer beyond educational diagnosis is limited.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Narrow educational application",
        "Limited impact on how modern AI systems are built or served",
        "Benchmark/infrastructure value exceeds broader research novelty"
      ],
      "what_would_raise_priority": "Evidence that malrule-style executable misconception modeling improves general reasoning evaluation or tutoring performance in widely used frontier models would raise its priority.",
      "one_sentence_episode_hook": "LLMs can often get the math right\u2014but can they recognize the exact wrong rule a student is using and predict their next mistake?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03184",
      "title": "Decentralized Autoregressive Generation",
      "abstract": "We present a theoretical analysis of decentralization of autoregressive generation. We define the Decentralized Discrete Flow Matching objective, by expressing probability generating velocity as a linear combination of expert flows. We also conduct experiments demonstrating the equivalence between decentralized and centralized training settings for multimodal language models across diverse set of benchmarks. Specifically, we compare two distinct paradigms: LLaVA and InternVL 2.5-1B, which uses a fixed CLIP vision encoder and performs full-parameter fine-tuning (ViT+MLP+LLM) during the instruction tuning stage.",
      "authors": [
        "Stepan Maschan",
        "Haoxuan Qu",
        "Jun Liu"
      ],
      "published_at": "2026-01-06T17:07:27",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03184v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.504040002822876,
      "sim_memory": 0.4785934388637543,
      "sim_negative": 0.3507066071033478,
      "broad_relevance": 0.504040002822876,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.49964356422424316,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.4785934388637543,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.5499999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3561585354804993,
      "memory_score": 0.019578031659126277,
      "quality_score": 0.525,
      "bridge_score": 0.019578031659126277,
      "max_axis_score": 0.3561585354804993,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory",
        "Training",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "Decentralizing autoregressive generation is an interesting theoretical angle, and the paper tries to connect that theory to multimodal LLM training setups like LLaVA and InternVL. If decentralized and centralized training are truly equivalent in practice, that could matter for future distributed model design.",
      "why_not_higher": "The memory/storage relevance is weak: this is not really about KV cache, offload, bandwidth, paging, or serving systems. The evidence also appears limited to equivalence demonstrations on a small set of multimodal models rather than a clear change in how practitioners build or serve mainstream AI systems.",
      "downgrade_reasons": [
        "memory connection is mostly absent",
        "theory-heavy with unclear systems payoff",
        "limited evidence on realistic large-scale deployment settings",
        "reads more like equivalence/benchmarking than a practical new recipe",
        "broad-audience urgency is modest"
      ],
      "what_would_raise_priority": "Show concrete training or serving wins at scale\u2014such as lower communication cost, better distributed efficiency, or simpler deployment on large multimodal or language models.",
      "one_sentence_episode_hook": "Can autoregressive generation be split across decentralized experts without losing anything\u2014and does that actually help real multimodal LLM systems?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03121",
      "title": "ToxiGAN: Toxic Data Augmentation via LLM-Guided Directional Adversarial Generation",
      "abstract": "Augmenting toxic language data in a controllable and class-specific manner is crucial for improving robustness in toxicity classification, yet remains challenging due to limited supervision and distributional skew. We propose ToxiGAN, a class-aware text augmentation framework that combines adversarial generation with semantic guidance from large language models (LLMs). To address common issues in GAN-based augmentation such as mode collapse and semantic drift, ToxiGAN introduces a two-step directional training strategy and leverages LLM-generated neutral texts as semantic ballast. Unlike prior work that treats LLMs as static generators, our approach dynamically selects neutral exemplars to provide balanced guidance. Toxic samples are explicitly optimized to diverge from these exemplars, reinforcing class-specific contrastive signals. Experiments on four hate speech benchmarks show that ToxiGAN achieves the strongest average performance in both macro-F1 and hate-F1, consistently outperforming traditional and LLM-based augmentation methods. Ablation and sensitivity analyses further confirm the benefits of semantic ballast and directional training in enhancing classifier robustness.",
      "authors": [
        "Peiran Li",
        "Jan Fillies",
        "Adrian Paschke"
      ],
      "published_at": "2026-01-06T15:50:46",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03121v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.3619881868362427,
      "sim_memory": 0.37047600746154785,
      "sim_negative": 0.19350585341453552,
      "broad_relevance": 0.3619881868362427,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6502503752708435,
      "evidence_score": 0.43000000000000005,
      "direct_memory_relevance": 0.37047600746154785,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.35613401234149933,
      "memory_score": 0.0,
      "quality_score": 0.595,
      "bridge_score": 0.0,
      "max_axis_score": 0.35613401234149933,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "Synthetic data augmentation for toxic language remains a live issue as teams try to harden safety classifiers and reduce class imbalance with LLM-assisted pipelines. The paper also reflects a broader trend of using LLMs as guidance modules rather than end-to-end generators.",
      "why_not_higher": "This is still a narrow toxicity-classification augmentation paper, not a result that clearly changes mainstream LLM building or serving practice. Its memory/storage relevance is essentially absent, and the claimed gains are benchmark-level rather than systems-shaping.",
      "downgrade_reasons": [
        "narrow task focus",
        "no direct memory/storage mechanism",
        "limited implications for modern AI system design",
        "primarily benchmark improvement"
      ],
      "what_would_raise_priority": "Show that the directional LLM-guided augmentation framework transfers across multiple safety, moderation, and general low-resource text classification settings at production scale.",
      "one_sentence_episode_hook": "Can an LLM act as semantic ballast for a GAN to generate better toxic data\u2014or is ToxiGAN just a clever benchmark win in a narrow moderation niche?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03159",
      "title": "Rapid Augmentations for Time Series (RATS): A High-Performance Library for Time Series Augmentation",
      "abstract": "Time series augmentation is critical for training robust deep learning models, particularly in domains where labelled data is scarce and expensive to obtain. However, existing augmentation libraries for time series, mainly written in Python, suffer from performance bottlenecks, where running time grows exponentially as dataset sizes increase -- an aspect limiting their applicability in large-scale, production-grade systems. We introduce RATS (Rapid Augmentations for Time Series), a high-performance library for time series augmentation written in Rust with Python bindings (RATSpy). RATS implements multiple augmentation methods spanning basic transformations, frequency-domain operations and time warping techniques, all accessible through a unified pipeline interface with built-in parallelisation. Comprehensive benchmarking of RATSpy versus a commonly used library (tasug) on 143 datasets demonstrates that RATSpy achieves an average speedup of 74.5\\% over tsaug (up to 94.8\\% on large datasets), with up to 47.9\\% less peak memory usage.",
      "authors": [
        "Wadie Skaf",
        "Felix Kern",
        "Aryamaan Basu Roy",
        "Tejas Pradhan",
        "Roman Kalkreuth",
        "Holger Hoos"
      ],
      "published_at": "2026-01-06T16:33:51",
      "categories": [
        "cs.LG",
        "cs.AI",
        "cs.PF"
      ],
      "url": "http://arxiv.org/abs/2601.03159v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4614270329475403,
      "sim_memory": 0.3920440673828125,
      "sim_negative": 0.42989909648895264,
      "broad_relevance": 0.4614270329475403,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5830981135368347,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.3920440673828125,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.54,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3558928269147873,
      "memory_score": 0.14761322021484374,
      "quality_score": 0.525,
      "bridge_score": 0.14761322021484374,
      "max_axis_score": 0.3558928269147873,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Systems",
        "Training",
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "It targets a real production pain point in time-series ML: Python augmentation pipelines can become slow and memory-hungry at scale, and a Rust implementation with Python bindings is a practical engineering response. The benchmarking across many datasets gives it some immediate utility for practitioners working on data pipelines.",
      "why_not_higher": "This is mainly a domain-specific tooling and benchmark paper rather than a result that changes mainstream AI model building or serving. Its memory angle is secondary peak-memory reduction in augmentation workloads, not a first-class AI memory/cache/storage mechanism.",
      "downgrade_reasons": [
        "narrow time-series scope",
        "memory connection is adjacent, not core",
        "primarily library engineering plus benchmarks",
        "limited relevance to modern foundation-model workflows",
        "does not clearly alter training or inference practice outside its niche"
      ],
      "what_would_raise_priority": "A stronger systems story showing impact on large-scale multimodal or foundation-model data pipelines, with clearer links to data movement, bandwidth, and end-to-end training throughput, would raise priority.",
      "one_sentence_episode_hook": "Can rewriting a time-series augmentation stack in Rust turn data preprocessing from a hidden bottleneck into a production-ready systems advantage?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03213",
      "title": "Critic-Guided Reinforcement Unlearning in Text-to-Image Diffusion",
      "abstract": "Machine unlearning in text-to-image diffusion models aims to remove targeted concepts while preserving overall utility. Prior diffusion unlearning methods typically rely on supervised weight edits or global penalties; reinforcement-learning (RL) approaches, while flexible, often optimize sparse end-of-trajectory rewards, yielding high-variance updates and weak credit assignment. We present a general RL framework for diffusion unlearning that treats denoising as a sequential decision process and introduces a timestep-aware critic with noisy-step rewards. Concretely, we train a CLIP-based reward predictor on noisy latents and use its per-step signal to compute advantage estimates for policy-gradient updates of the reverse diffusion kernel. Our algorithm is simple to implement, supports off-policy reuse, and plugs into standard text-to-image backbones. Across multiple concepts, the method achieves better or comparable forgetting to strong baselines while maintaining image quality and benign prompt fidelity; ablations show that (i) per-step critics and (ii) noisy-conditioned rewards are key to stability and effectiveness. We release code and evaluation scripts to facilitate reproducibility and future research on RL-based diffusion unlearning.",
      "authors": [
        "Mykola Vysotskyi",
        "Zahar Kohut",
        "Mariia Shpir",
        "Taras Rumezhak",
        "Volodymyr Karpiv"
      ],
      "published_at": "2026-01-06T17:52:02",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03213v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "multimodal",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.37604790925979614,
      "sim_memory": 0.4343532621860504,
      "sim_negative": 0.3010849952697754,
      "broad_relevance": 0.37604790925979614,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.4860326051712036,
      "evidence_score": 0.35000000000000003,
      "direct_memory_relevance": 0.4343532621860504,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7699999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3557192635536194,
      "memory_score": 0.060305978655815135,
      "quality_score": 0.575,
      "bridge_score": 0.060305978655815135,
      "max_axis_score": 0.3557192635536194,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Training",
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "Diffusion unlearning is timely as model providers look for practical ways to remove unwanted concepts without full retraining. The paper also reframes diffusion denoising as an RL problem with denser per-timestep credit assignment, which is a potentially useful training idea.",
      "why_not_higher": "This is still a fairly narrow text-to-image unlearning paper rather than a result that broadly changes how modern AI systems are built or served. Its memory/storage relevance is essentially absent, and the evidence appears centered on benchmark-style concept forgetting rather than deployment-scale workloads or systems outcomes.",
      "downgrade_reasons": [
        "Narrow task focus on text-to-image concept unlearning",
        "Memory/storage connection absent beyond superficial training mechanics",
        "Unclear impact on mainstream model training or serving practice",
        "Evidence appears limited to unlearning benchmarks rather than realistic production settings",
        "Episode risks becoming a specialized benchmark discussion"
      ],
      "what_would_raise_priority": "Show that critic-guided per-step rewards generalize beyond diffusion unlearning to broadly useful training or safety editing workflows at meaningful model scale.",
      "one_sentence_episode_hook": "Can reinforcement learning make diffusion unlearning actually stable and precise, or is this just a better way to chase a niche benchmark?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03471",
      "title": "EpiQAL: Benchmarking Large Language Models in Epidemiological Question Answering for Enhanced Alignment and Reasoning",
      "abstract": "Reliable epidemiological reasoning requires synthesizing study evidence to infer disease burden, transmission dynamics, and intervention effects at the population level. Existing medical question answering benchmarks primarily emphasize clinical knowledge or patient-level reasoning, yet few systematically evaluate evidence-grounded epidemiological inference. We present EpiQAL, the first diagnostic benchmark for epidemiological question answering across diverse diseases, comprising three subsets built from open-access literature. The three subsets progressively test factual recall, multi-step inference, and conclusion reconstruction under incomplete information, and are constructed through a quality-controlled pipeline combining taxonomy guidance, multi-model verification, and difficulty screening. Experiments on fourteen models spanning open-source and proprietary systems reveal that current LLMs show limited performance on epidemiological reasoning, with multi-step inference posing the greatest challenge. Model rankings shift across subsets, and scale alone does not predict success. Chain-of-Thought prompting benefits multi-step inference but yields mixed results elsewhere. EpiQAL provides fine-grained diagnostic signals for evidence-grounding, inferential reasoning, and conclusion reconstruction.",
      "authors": [
        "Mingyang Wei",
        "Dehai Min",
        "Zewen Liu",
        "Yuzhang Xie",
        "Guanchen Wu",
        "Ziyang Zhang",
        "Carl Yang",
        "Max S. Y. Lau",
        "Qi He",
        "Lu Cheng"
      ],
      "published_at": "2026-01-06T23:49:10",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03471v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4374869465827942,
      "sim_memory": 0.20630866289138794,
      "sim_negative": 0.17755071818828583,
      "broad_relevance": 0.4374869465827942,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6286912858486176,
      "evidence_score": 0.46,
      "direct_memory_relevance": 0.20630866289138794,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.48,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3555497768521309,
      "memory_score": 0.011892598867416393,
      "quality_score": 0.525,
      "bridge_score": 0.011892598867416393,
      "max_axis_score": 0.3555497768521309,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "It highlights a real weakness in current LLMs: evidence-grounded epidemiological reasoning is still brittle, especially for multi-step inference. As domain benchmarks proliferate, this is a timely reminder that scale and generic QA success do not guarantee reliable scientific reasoning.",
      "why_not_higher": "The contribution is mainly a domain-specific benchmark rather than a method that changes how mainstream AI systems are built or served. Its connection to the podcast's memory/storage lens is essentially absent, and its transfer beyond epidemiology is limited.",
      "downgrade_reasons": [
        "narrow domain benchmark",
        "no direct memory/storage mechanism",
        "limited systems relevance",
        "mostly evaluative rather than methodological",
        "unclear broad audience pull beyond specialized healthcare/biomed listeners"
      ],
      "what_would_raise_priority": "Evidence that the benchmark exposes general failure modes across scientific QA and directly motivates new training, retrieval, or inference methods with broad applicability would raise priority.",
      "one_sentence_episode_hook": "What happens when you ask top LLMs to reason like epidemiologists rather than autocomplete textbooks\u2014and they still stumble on evidence synthesis and multi-step public-health inference?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03195",
      "title": "Sparse Knowledge Distillation: A Mathematical Framework for Probability-Domain Temperature Scaling and Multi-Stage Compression",
      "abstract": "We develop a unified theoretical framework for sparse knowledge distillation based on probability-domain softening operators. While the equivalence $p^{1/T} \\propto \\mathrm{softmax}(z/T)$ is well known, our contribution is an operator-level analytical framework built on this foundation rather than the equivalence itself.   The framework comprises four core components: (i) operator-agnostic bias--variance decompositions that characterize when sparse students outperform dense teachers, (ii) a homotopy path formalization of multi-stage pruning in function space explaining why iterative compression succeeds where one-shot pruning fails, (iii) convergence guarantees establishing $O(1/n)$ rates for $n$-stage distillation with explicit parameter dependence, and (iv) equivalence class characterizations identifying distinct probability-domain operators that yield identical student models under capacity constraints.   We introduce an axiomatic definition of probability-domain softening operators based on ranking preservation, continuity, entropy monotonicity, identity, and boundary behavior, and show that multiple non-equivalent operator families satisfy these axioms. All learning-theoretic guarantees are shown to hold uniformly across this operator class, independent of implementation details. These results provide theoretical grounding for black-box teacher distillation, partial-access settings such as top-$k$ truncation and text-only outputs, and privacy-preserving model compression.",
      "authors": [
        "Aaron R. Flouro",
        "Shawn P. Chadwick"
      ],
      "published_at": "2026-01-06T17:17:24",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03195v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 5,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "theory",
      "narrow_domain_flag": false,
      "sim_public": 0.36616939306259155,
      "sim_memory": 0.3764804005622864,
      "sim_negative": 0.24419495463371277,
      "broad_relevance": 0.36616939306259155,
      "momentum": 0.06108604879161034,
      "teachability": 0.75,
      "novelty_score": 0.5927731394767761,
      "evidence_score": 0.3191759469228055,
      "direct_memory_relevance": 0.3764804005622864,
      "systems_leverage": 0.4,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7599999999999999,
      "clarity": 0.6,
      "reproducibility": 0.3,
      "public_interest_score": 0.3528603906370368,
      "memory_score": 0.09244523379942254,
      "quality_score": 0.5566703787691222,
      "bridge_score": 0.09244523379942254,
      "max_axis_score": 0.3528603906370368,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "Model compression and black-box distillation remain relevant, and this paper tries to give a unified mathematical account of sparse distillation, top-k outputs, and multi-stage pruning. That could be useful background if distillation-driven deployment becomes a near-term editorial theme.",
      "why_not_higher": "The contribution appears primarily theoretical, with no clear evidence on realistic LLMs, serving stacks, or hardware-constrained workloads. Its memory relevance is indirect: compression is discussed, but not through concrete KV-cache, bandwidth, offload, paging, or locality mechanisms.",
      "downgrade_reasons": [
        "primarily theory with limited systems evidence",
        "memory/storage link is adjacent rather than direct",
        "no demonstrated impact on modern large-model build or serve practice",
        "unclear validation on realistic workloads or hardware"
      ],
      "what_would_raise_priority": "Show strong empirical results on modern LLM distillation or pruning under real memory/bandwidth constraints, with measurable serving or deployment gains.",
      "one_sentence_episode_hook": "Can a clean mathematical theory explain why iterative sparse distillation works better than one-shot compression\u2014and when a smaller student can actually beat its teacher?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03025",
      "title": "LittiChoQA: Literary Texts in Indic Languages Chosen for Question Answering",
      "abstract": "Long-context question answering (QA) over literary texts poses significant challenges for modern large language models, particularly in low-resource languages. We address the scarcity of long-context QA resources for Indic languages by introducing LittiChoQA, the largest literary QA dataset to date covering many languages spoken in the Gangetic plains of India. The dataset comprises over 270K automatically generated question-answer pairs with a balanced distribution of factoid and non-factoid questions, generated from naturally authored literary texts collected from the open web. We evaluate multiple multilingual LLMs on non-factoid, abstractive QA, under both full-context and context-shortened settings. Results demonstrate a clear trade-off between performance and efficiency: full-context fine-tuning yields the highest token-level and semantic-level scores, while context shortening substantially improves throughput. Among the evaluated models, Krutrim-2 achieves the strongest performance, obtaining a semantic score of 76.1 with full context. While, in shortened context settings it scores 74.9 with answer paragraph selection and 71.4 with vector-based retrieval. Qualitative evaluations further corroborate these findings.",
      "authors": [
        "Aarya Khandelwal",
        "Ritwik Mishra",
        "Rajiv Ratn Shah"
      ],
      "published_at": "2026-01-06T13:59:41",
      "categories": [
        "cs.CL"
      ],
      "url": "http://arxiv.org/abs/2601.03025v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4466358423233032,
      "sim_memory": 0.3593677282333374,
      "sim_negative": 0.14124253392219543,
      "broad_relevance": 0.4466358423233032,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5891541242599487,
      "evidence_score": 0.38,
      "direct_memory_relevance": 0.3593677282333374,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.48,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.35236387133598324,
      "memory_score": 0.0238103184700012,
      "quality_score": 0.525,
      "bridge_score": 0.0238103184700012,
      "max_axis_score": 0.35236387133598324,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "It addresses a real gap in long-context QA resources for low-resource Indic languages and gives timely empirical evidence on the accuracy-versus-efficiency tradeoff between full-context and retrieval-based shortening. That makes it relevant to multilingual evaluation and dataset building right now.",
      "why_not_higher": "This is primarily a dataset-and-benchmark paper for a narrow task domain rather than a paper that changes mainstream model building or serving practice. Its efficiency discussion is incidental to QA setup choices, not a concrete memory/storage systems contribution.",
      "downgrade_reasons": [
        "Narrow literary QA benchmark",
        "Memory connection is indirect",
        "Limited systems novelty",
        "Unclear impact beyond Indic long-context evaluation",
        "Mostly comparative results rather than a new method"
      ],
      "what_would_raise_priority": "A stronger cross-lingual retrieval/serving method with clear implications for long-context inference efficiency across domains would raise this substantially.",
      "one_sentence_episode_hook": "What do you gain\u2014and lose\u2014when you shrink long literary context for multilingual QA in low-resource Indic languages?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.08857",
      "title": "Revisiting Software Engineering Education in the Era of Large Language Models: A Curriculum Adaptation and Academic Integrity Framework",
      "abstract": "The integration of Large Language Models (LLMs), such as ChatGPT and GitHub Copilot, into professional workflows is increasingly reshaping software engineering practices. These tools have lowered the cost of code generation, explanation, and testing, while introducing new forms of automation into routine development tasks. In contrast, most of the software engineering and computer engineering curricula remain closely aligned with pedagogical models that equate manual syntax production with technical competence. This growing misalignment raises concerns regarding assessment validity, learning outcomes, and the development of foundational skills. Adopting a conceptual research approach, this paper proposes a theoretical framework for analyzing how generative AI alters core software engineering competencies and introduces a pedagogical design model for LLM-integrated education. Attention is given to computer engineering programs in Turkey, where centralized regulation, large class sizes, and exam-oriented assessment practices amplify these challenges. The framework delineates how problem analysis, design, implementation, and testing increasingly shift from construction toward critique, validation, and human-AI stewardship. In addition, the paper argues that traditional plagiarism-centric integrity mechanisms are becoming insufficient, motivating a transition toward a process transparency model. While this work provides a structured proposal for curriculum adaptation, it remains a theoretical contribution; the paper concludes by outlining the need for longitudinal empirical studies to evaluate these interventions and their long-term impacts on learning.",
      "authors": [
        "Mustafa Degerli"
      ],
      "published_at": "2026-01-06T22:41:47",
      "categories": [
        "cs.SE",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.08857v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.46233418583869934,
      "sim_memory": 0.18610720336437225,
      "sim_negative": 0.12561671435832977,
      "broad_relevance": 0.46233418583869934,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5535430908203125,
      "evidence_score": 0.2,
      "direct_memory_relevance": 0.18610720336437225,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7499999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3517317193746567,
      "memory_score": 0.0018321610093116891,
      "quality_score": 0.575,
      "bridge_score": 0.0018321610093116891,
      "max_axis_score": 0.3517317193746567,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "LLMs are already changing programming education and assessment, so a curriculum and integrity framework is timely for universities deciding how to respond this academic year. It also speaks to a broad anxiety in AI adoption: when automation changes what competence should mean.",
      "why_not_higher": "This is a theoretical education-policy paper rather than a paper that changes AI system design, training, inference, or evaluation practice. It has no empirical validation on realistic models, deployments, classrooms, or longitudinal outcomes yet.",
      "downgrade_reasons": [
        "No direct memory/storage contribution",
        "Primarily pedagogy and academic integrity, not AI systems",
        "Conceptual framework without empirical results",
        "Limited relevance to how practitioners build or serve models",
        "Broad topic but weak novelty signal for the podcast's recent research focus"
      ],
      "what_would_raise_priority": "A strong empirical follow-up with controlled classroom studies, concrete assessment redesigns, and measurable effects on learning and integrity would raise its priority.",
      "one_sentence_episode_hook": "If Copilot and ChatGPT now do the syntax, what exactly should software engineering education still be teaching humans?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03018",
      "title": "Dementia-R1: Reinforced Pretraining and Reasoning from Unstructured Clinical Notes for Real-World Dementia Prognosis",
      "abstract": "While Large Language Models (LLMs) have shown strong performance on clinical text understanding, they struggle with longitudinal prediction tasks such as dementia prognosis, which require reasoning over complex, non-monotonic symptom trajectories across multiple visits. Standard supervised training lacks explicit annotations for symptom evolution, while direct Reinforcement Learning (RL) is hindered by sparse binary rewards. To address this challenge, we introduce Dementia-R1, an RL-based framework for longitudinal dementia prognosis from unstructured clinical notes. Our approach adopts a Cold-Start RL strategy that pre-trains the model to predict verifiable clinical indices extracted from patient histories, enhancing the capability to reason about disease progression before determining the final clinical status. Extensive experiments show that Dementia-R1 achieves the best overall performance on the AMC real-world unstructured cohort, reaching an AUROC of 84.02% and outperforming models up to 10x larger. The framework also generalizes to Parkinson's disease dementia prediction in an independent hospital cohort, achieving an AUROC of 78.37%. On the ADNI benchmark, our 7B model attains the highest AUROC among all LLM baselines at 83.17%, demonstrating strong longitudinal reasoning over fluctuating cognitive trajectories. Code is available at https://anonymous.4open.science/r/dementiar1-CDB5.",
      "authors": [
        "Choonghan Kim",
        "Hyunmin Hwang",
        "Hangeol Chang",
        "Jaemin Kim",
        "Jinse Park",
        "Jae-Sung Lim",
        "Jong Chul Ye"
      ],
      "published_at": "2026-01-06T13:44:04",
      "categories": [
        "cs.CL",
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03018v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.44378840923309326,
      "sim_memory": 0.3856459856033325,
      "sim_negative": 0.32861900329589844,
      "broad_relevance": 0.44378840923309326,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5705558061599731,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.3856459856033325,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3487198936939239,
      "memory_score": 0.0,
      "quality_score": 0.575,
      "bridge_score": 0.0,
      "max_axis_score": 0.3487198936939239,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "It shows a practical way to use reinforced pretraining on clinical notes for longitudinal prognosis, a setting where standard supervised fine-tuning struggles. The reported gains over larger models and cross-cohort validation make it more credible than a purely single-dataset benchmark.",
      "why_not_higher": "This is still primarily a narrow clinical application paper rather than a result that clearly changes mainstream AI practice. Its memory/storage relevance is effectively absent, and the RL setup appears specialized to prognosis with verifiable intermediate clinical signals.",
      "downgrade_reasons": [
        "narrow medical application",
        "limited transfer beyond clinical prognosis",
        "not a memory/storage paper",
        "benchmark-heavy framing",
        "unclear impact on mainstream LLM training or serving"
      ],
      "what_would_raise_priority": "Stronger evidence that the cold-start RL recipe transfers cleanly to other longitudinal reasoning tasks outside medicine would make it much more episode-worthy.",
      "one_sentence_episode_hook": "Can reinforcement learning teach a smaller LLM to reason over messy, years-long clinical histories better than models 10 times larger?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03054",
      "title": "IBISAgent: Reinforcing Pixel-Level Visual Reasoning in MLLMs for Universal Biomedical Object Referring and Segmentation",
      "abstract": "Recent research on medical MLLMs has gradually shifted its focus from image-level understanding to fine-grained, pixel-level comprehension. Although segmentation serves as the foundation for pixel-level understanding, existing approaches face two major challenges. First, they introduce implicit segmentation tokens and require simultaneous fine-tuning of both the MLLM and external pixel decoders, which increases the risk of catastrophic forgetting and limits generalization to out-of-domain scenarios. Second, most methods rely on single-pass reasoning and lack the capability to iteratively refine segmentation results, leading to suboptimal performance. To overcome these limitations, we propose a novel agentic MLLM, named IBISAgent, that reformulates segmentation as a vision-centric, multi-step decision-making process. IBISAgent enables MLLMs to generate interleaved reasoning and text-based click actions, invoke segmentation tools, and produce high-quality masks without architectural modifications. By iteratively performing multi-step visual reasoning on masked image features, IBISAgent naturally supports mask refinement and promotes the development of pixel-level visual reasoning capabilities. We further design a two-stage training framework consisting of cold-start supervised fine-tuning and agentic reinforcement learning with tailored, fine-grained rewards, enhancing the model's robustness in complex medical referring and reasoning segmentation tasks. Extensive experiments demonstrate that IBISAgent consistently outperforms both closed-source and open-source SOTA methods. All datasets, code, and trained models will be released publicly.",
      "authors": [
        "Yankai Jiang",
        "Qiaoru Li",
        "Binlu Xu",
        "Haoran Sun",
        "Chao Ding",
        "Junting Dong",
        "Yuxiang Cai",
        "Xuhong Zhang",
        "Jianwei Yin"
      ],
      "published_at": "2026-01-06T14:37:50",
      "categories": [
        "cs.CV",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03054v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 2,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "vision",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4115166962146759,
      "sim_memory": 0.2899097204208374,
      "sim_negative": 0.4316073954105377,
      "broad_relevance": 0.4115166962146759,
      "momentum": 0.031333575283049045,
      "teachability": 0.7,
      "novelty_score": 0.533419281244278,
      "evidence_score": 0.48986122886681094,
      "direct_memory_relevance": 0.2899097204208374,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.48,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3462138004376759,
      "memory_score": 0.01615626359026856,
      "quality_score": 0.5889444915467243,
      "bridge_score": 0.01615626359026856,
      "max_axis_score": 0.3462138004376759,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training",
        "Inference"
      ],
      "status": "Deferred this cycle",
      "why_now": "It reflects a current trend toward agentic, iterative visual reasoning in multimodal models and applies that idea to biomedical segmentation without adding a custom decoder. The combination of supervised fine-tuning plus RL for text-driven click refinement is timely within medical MLLM work.",
      "why_not_higher": "This is still primarily a domain-specific biomedical segmentation paper, and the core contribution does not clearly change mainstream AI system design outside that niche. Its memory/storage relevance is effectively absent, and the broader transferability beyond medical referring segmentation is not yet established.",
      "downgrade_reasons": [
        "narrow biomedical task focus",
        "memory/storage connection absent",
        "limited evidence of transfer beyond medical segmentation",
        "benchmark-oriented application framing"
      ],
      "what_would_raise_priority": "Show strong transfer of the same agentic pixel-reasoning setup to general-purpose multimodal models and non-medical segmentation or grounding tasks at meaningful scale.",
      "one_sentence_episode_hook": "Can a multimodal model learn to segment by thinking out loud and clicking its way to a better mask, instead of relying on a bespoke decoder?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03056",
      "title": "Fine-Grained Generalization via Structuralizing Concept and Feature Space into Commonality, Specificity and Confounding",
      "abstract": "Fine-Grained Domain Generalization (FGDG) presents greater challenges than conventional domain generalization due to the subtle inter-class differences and relatively pronounced intra-class variations inherent in fine-grained recognition tasks. Under domain shifts, the model becomes overly sensitive to fine-grained cues, leading to the suppression of critical features and a significant drop in performance. Cognitive studies suggest that humans classify objects by leveraging both common and specific attributes, enabling accurate differentiation between fine-grained categories. However, current deep learning models have yet to incorporate this mechanism effectively. Inspired by this mechanism, we propose Concept-Feature Structuralized Generalization (CFSG). This model explicitly disentangles both the concept and feature spaces into three structured components: common, specific, and confounding segments. To mitigate the adverse effects of varying degrees of distribution shift, we introduce an adaptive mechanism that dynamically adjusts the proportions of common, specific, and confounding components. In the final prediction, explicit weights are assigned to each pair of components. Extensive experiments on three single-source benchmark datasets demonstrate that CFSG achieves an average performance improvement of 9.87% over baseline models and outperforms existing state-of-the-art methods by an average of 3.08%. Additionally, explainability analysis validates that CFSG effectively integrates multi-granularity structured knowledge and confirms that feature structuralization facilitates the emergence of concept structuralization.",
      "authors": [
        "Zhen Wang",
        "Jiaojiao Zhao",
        "Qilong Wang",
        "Yongfeng Dong",
        "Wenlong Yu"
      ],
      "published_at": "2026-01-06T14:39:09",
      "categories": [
        "cs.CV"
      ],
      "url": "http://arxiv.org/abs/2601.03056v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "vision",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.4061078131198883,
      "sim_memory": 0.3768928349018097,
      "sim_negative": 0.3450204133987427,
      "broad_relevance": 0.4061078131198883,
      "momentum": 0.017490950964499406,
      "teachability": 0.7,
      "novelty_score": 0.5943847298622131,
      "evidence_score": 0.40931471805599456,
      "direct_memory_relevance": 0.3768928349018097,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.52,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3448854513165975,
      "memory_score": 0.0,
      "quality_score": 0.5727258872223978,
      "bridge_score": 0.0,
      "max_axis_score": 0.3448854513165975,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "Domain generalization for fine-grained vision is a real problem, and the paper offers an interpretable structured decomposition into common, specific, and confounding factors. That could interest listeners who follow robustness and representation learning in vision.",
      "why_not_higher": "This is still a narrow fine-grained CV benchmark paper rather than a result that clearly changes mainstream AI system design or evaluation. The memory/storage connection is absent, and the evidence is limited to three single-source benchmark datasets rather than broader modern workloads.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Narrow task focus in fine-grained vision",
        "Benchmark-driven gains without clear systems impact",
        "Limited evidence on realistic large-scale models or deployments",
        "Unclear transfer beyond the stated domain-generalization setting"
      ],
      "what_would_raise_priority": "Show strong transfer to broader multimodal or foundation-model settings, with evidence that the structured decomposition improves robustness across widely used real-world benchmarks.",
      "one_sentence_episode_hook": "Can AI learn to separate what objects share, what makes them unique, and what is just confounding noise\u2014and does that finally help fine-grained models survive domain shift?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03038",
      "title": "Validating Generalist Robots with Situation Calculus and STL Falsification",
      "abstract": "Generalist robots are becoming a reality, capable of interpreting natural language instructions and executing diverse operations. However, their validation remains challenging because each task induces its own operational context and correctness specification, exceeding the assumptions of traditional validation methods. We propose a two-layer validation framework that combines abstract reasoning with concrete system falsification. At the abstract layer, situation calculus models the world and derives weakest preconditions, enabling constraint-aware combinatorial testing to systematically generate diverse, semantically valid world-task configurations with controllable coverage strength. At the concrete layer, these configurations are instantiated for simulation-based falsification with STL monitoring. Experiments on tabletop manipulation tasks show that our framework effectively uncovers failure cases in the NVIDIA GR00T controller, demonstrating its promise for validating general-purpose robot autonomy.",
      "authors": [
        "Changwen Li",
        "Rongjie Yan",
        "Chih-Hong Cheng",
        "Jian Zhang"
      ],
      "published_at": "2026-01-06T14:13:33",
      "categories": [
        "cs.RO",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03038v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "application",
      "domain_bucket": "robotics",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4481405019760132,
      "sim_memory": 0.25116974115371704,
      "sim_negative": 0.25524431467056274,
      "broad_relevance": 0.4481405019760132,
      "momentum": 0.0,
      "teachability": 0.7,
      "novelty_score": 0.5881061255931854,
      "evidence_score": 0.37,
      "direct_memory_relevance": 0.25116974115371704,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.54,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3426580694317818,
      "memory_score": 0.01835092234611513,
      "quality_score": 0.545,
      "bridge_score": 0.01835092234611513,
      "max_axis_score": 0.3426580694317818,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Systems"
      ],
      "status": "Deferred this cycle",
      "why_now": "Generalist robot controllers are arriving faster than robust validation methods, and this paper offers a structured way to generate semantically valid test scenarios and falsify failures. That makes it timely for people thinking about how to evaluate embodied foundation models beyond ad hoc demos.",
      "why_not_higher": "The contribution is mainly a robotics validation workflow, not something that broadly changes how most AI practitioners build or serve modern models. It also has no meaningful memory/storage angle under the show's second lens.",
      "downgrade_reasons": [
        "robotics-specific validation paper",
        "weak fit to memory/storage lens",
        "limited evidence beyond tabletop simulation and one controller",
        "evaluation methodology is useful but not broadly transformative for mainstream AI"
      ],
      "what_would_raise_priority": "Stronger cross-platform evidence on multiple embodied foundation models and real-robot deployments, plus clearer transfer to general AI evaluation practice, would raise priority.",
      "one_sentence_episode_hook": "If generalist robots can follow your instructions, how do you systematically discover the weird edge cases where they fail before they touch the real world?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03453",
      "title": "Measures of classification bias derived from sample size analysis",
      "abstract": "We propose the use of a simple intuitive principle for measuring algorithmic classification bias: the significance of the differences in a classifier's error rates across the various demographics is inversely commensurate with the sample size required to statistically detect them. That is, if large sample sizes are required to statistically establish biased behavior, the algorithm is less biased, and vice versa. In a simple setting, we assume two distinct demographics, and non-parametric estimates of the error rates on them, e1 and e2, respectively. We use a well-known approximate formula for the sample size of the chi-squared test, and verify some basic desirable properties of the proposed measure. Next, we compare the proposed measure with two other commonly used statistics, the difference e2-e1 and the ratio e2/e1 of the error rates. We establish that the proposed measure is essentially different in that it can rank algorithms for bias differently, and we discuss some of its advantages over the other two measures. Finally, we briefly discuss how some of the desirable properties of the proposed measure emanate from fundamental characteristics of the method, rather than the approximate sample size formula we used, and thus, are expected to hold in more complex settings with more than two demographics.",
      "authors": [
        "Ioannis Ivrissimtzis",
        "Shauna Concannon",
        "Matthew Houliston",
        "Graham Roberts"
      ],
      "published_at": "2026-01-06T22:50:56",
      "categories": [
        "stat.ME",
        "cs.CY",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03453v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.2839810252189636,
      "sim_memory": 0.19892822206020355,
      "sim_negative": 0.1720278561115265,
      "broad_relevance": 0.2839810252189636,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6837685704231262,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.19892822206020355,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.5599999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.33775959312915804,
      "memory_score": 0.0,
      "quality_score": 0.545,
      "bridge_score": 0.0,
      "max_axis_score": 0.33775959312915804,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory"
      ],
      "status": "Deferred this cycle",
      "why_now": "Bias evaluation remains important as AI systems are deployed more widely, and the paper offers a different way to quantify classification disparity via sample-size-to-detect. It could interest a subset of responsible-AI and evaluation researchers.",
      "why_not_higher": "The contribution is mainly a statistical framing for fairness measurement, not something that clearly changes how modern foundation models are built, served, or evaluated in current practice. It also has no meaningful memory/storage relevance and appears light on realistic large-scale AI workloads.",
      "downgrade_reasons": [
        "No direct memory/storage connection",
        "Limited relevance to transformer or foundation-model practice",
        "Primarily methodological/statistical rather than systems-impacting",
        "Evidence appears to come from a simple two-demographic setting",
        "Unclear short-term pull for a broad AI audience"
      ],
      "what_would_raise_priority": "A stronger case on contemporary LLM or multimodal evaluation pipelines, with real deployment scenarios and comparisons showing decisions that materially change model auditing practice, would raise priority.",
      "one_sentence_episode_hook": "What if the fairest way to measure classification bias isn\u2019t the gap itself, but how many samples you need before the gap becomes statistically undeniable?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03237",
      "title": "PET-TURTLE: Deep Unsupervised Support Vector Machines for Imbalanced Data Clusters",
      "abstract": "Foundation vision, audio, and language models enable zero-shot performance on downstream tasks via their latent representations. Recently, unsupervised learning of data group structure with deep learning methods has gained popularity. TURTLE, a state of the art deep clustering algorithm, uncovers data labeling without supervision by alternating label and hyperplane updates, maximizing the hyperplane margin, in a similar fashion to support vector machines (SVMs). However, TURTLE assumes clusters are balanced; when data is imbalanced, it yields non-ideal hyperplanes that cause higher clustering error. We propose PET-TURTLE, which generalizes the cost function to handle imbalanced data distributions by a power law prior. Additionally, by introducing sparse logits in the labeling process, PET-TURTLE optimizes a simpler search space that in turn improves accuracy for balanced datasets. Experiments on synthetic and real data show that PET-TURTLE improves accuracy for imbalanced sources, prevents over-prediction of minority clusters, and enhances overall clustering.",
      "authors": [
        "Javier Salazar Cavazos"
      ],
      "published_at": "2026-01-06T18:30:25",
      "categories": [
        "cs.LG",
        "eess.IV",
        "stat.ML"
      ],
      "url": "http://arxiv.org/abs/2601.03237v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "systems",
      "narrow_domain_flag": false,
      "sim_public": 0.3736550211906433,
      "sim_memory": 0.31584563851356506,
      "sim_negative": 0.2594843804836273,
      "broad_relevance": 0.3736550211906433,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.6164608001708984,
      "evidence_score": 0.40931471805599456,
      "direct_memory_relevance": 0.31584563851356506,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3284610242841268,
      "memory_score": 0.0,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.0,
      "max_axis_score": 0.3284610242841268,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory"
      ],
      "status": "Deferred this cycle",
      "why_now": "It touches a timely theme\u2014using foundation-model latent spaces for unsupervised labeling\u2014and addresses a real weakness of prior deep clustering methods on imbalanced data. That could interest a niche ML audience working on clustering and representation learning.",
      "why_not_higher": "The contribution is mainly a clustering-objective refinement rather than something that changes mainstream LLM/vision system building or serving. The memory/storage link is absent, and the abstract does not show clear evaluation on large modern foundation-model workloads or production-relevant settings.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Narrow clustering-method paper",
        "Unclear impact on modern AI system design",
        "Evidence appears limited to synthetic and unspecified real datasets",
        "Likely low 30-90 day relevance for broad AI audience"
      ],
      "what_would_raise_priority": "Show strong gains on widely used foundation-model embedding benchmarks with clear downstream impact and a convincing path to large-scale practical deployment.",
      "one_sentence_episode_hook": "Can a margin-based deep clustering method finally stop over-predicting minority groups in imbalanced embedding spaces, or is PET-TURTLE still a niche benchmark story?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03442",
      "title": "Provable Acceleration of Distributed Optimization with Local Updates",
      "abstract": "In conventional distributed optimization, each agent performs a single local update between two communication rounds with its neighbors to synchronize solutions. Inspired by the success of using multiple local updates in federated learning, incorporating local updates into distributed optimization has recently attracted increasing attention. However, unlike federated learning, where multiple local updates can accelerate learning by improving gradient estimation under mini-batch settings, it remains unclear whether similar benefits hold in distributed optimization when gradients are exact. Moreover, existing theoretical results typically require reducing the step size when multiple local updates are employed, which can entirely offset any potential benefit of these additional local updates and obscure their true impact on convergence. In this paper, we focus on the classic DIGing algorithm and leverage the tight performance bounds provided by Performance Estimation Problems (PEP) to show that incorporating local updates can indeed accelerate distributed optimization. To the best of our knowledge, this is the first rigorous demonstration of such acceleration for a broad class of objective functions. Our analysis further reveals that, under an appropriate step size, performing only two local updates is sufficient to achieve the maximal possible improvement, and that additional local updates provide no further gains. Because more updates increase computational cost, these findings offer practical guidance for efficient implementation. Extensive experiments on both synthetic and real-world datasets corroborate the theoretical findings.",
      "authors": [
        "Zuang Wang",
        "Yongqiang Wang"
      ],
      "published_at": "2026-01-06T22:10:11",
      "categories": [
        "eess.SY",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03442v3",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "theory",
      "narrow_domain_flag": false,
      "sim_public": 0.5719354748725891,
      "sim_memory": 0.5309954285621643,
      "sim_negative": 0.22954343259334564,
      "broad_relevance": 0.5719354748725891,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5394402742385864,
      "evidence_score": 0.24,
      "direct_memory_relevance": 0.5309954285621643,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.52,
      "clarity": 0.6,
      "reproducibility": 0.3,
      "public_interest_score": 0.32749668359756473,
      "memory_score": 0.03029862856864929,
      "quality_score": 0.435,
      "bridge_score": 0.03029862856864929,
      "max_axis_score": 0.32749668359756473,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory",
        "Training"
      ],
      "status": "Deferred this cycle",
      "why_now": "It makes a clean theoretical claim that local updates can provably accelerate a classic distributed optimization method, with the practical takeaway that two local updates may be enough. That is timely for distributed training discussions where communication cost remains a bottleneck.",
      "why_not_higher": "The paper is optimization theory rather than a direct result on modern foundation-model training or serving, and its memory/storage relevance is weak. The evidence appears to be theory plus experiments on synthetic and real datasets, but not clearly on realistic large-model or hardware-constrained AI systems.",
      "downgrade_reasons": [
        "Distributed optimization theory without clear large-model systems impact",
        "Memory/storage connection is absent to weakly adjacent, not central",
        "No clear demonstration on transformer-scale training or serving workloads",
        "Likely limited 30-90 day interest for a broad AI audience"
      ],
      "what_would_raise_priority": "A follow-up showing communication-speedup tradeoffs on transformer or LLM training stacks, with concrete systems implications for distributed training efficiency, would raise priority.",
      "one_sentence_episode_hook": "Can just two local steps between communication rounds give distributed training a provable speedup\u2014or is this mostly a beautiful theory result looking for a modern AI use case?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03051",
      "title": "Temporal Graph Network: Hallucination Detection in Multi-Turn Conversation",
      "abstract": "Hallucinations can be produced by conversational AI systems, particularly in multi-turn conversations where context changes and contradictions may eventually surface. By representing the entire conversation as a temporal graph, we present a novel graph-based method for detecting dialogue-level hallucinations. Our framework models each dialogue as a node, encoding it using a sentence transformer. We explore two different ways of connectivity: i) shared-entity edges, which connect turns that refer to the same entities; ii) temporal edges, which connect contiguous turns in the conversation. Message-passing is used to update the node embeddings, allowing flow of information between related nodes. The context-aware node embeddings are then combined using attention pooling into a single vector, which is then passed on to a classifier to determine the presence and type of hallucinations. We demonstrate that our method offers slightly improved performance over existing methods. Further, we show the attention mechanism can be used to justify the decision making process. The code and model weights are made available at: https://github.com/sambuaneesh/anlp-project.",
      "authors": [
        "Vidhi Rathore",
        "Sambu Aneesh",
        "Himanshu Singh"
      ],
      "published_at": "2026-01-06T14:34:45",
      "categories": [
        "cs.CL",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03051v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.36915096640586853,
      "sim_memory": 0.26817312836647034,
      "sim_negative": 0.2846532464027405,
      "broad_relevance": 0.36915096640586853,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5927907824516296,
      "evidence_score": 0.30000000000000004,
      "direct_memory_relevance": 0.26817312836647034,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6599999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.309663907289505,
      "memory_score": 0.0,
      "quality_score": 0.575,
      "bridge_score": 0.0,
      "max_axis_score": 0.309663907289505,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "Hallucination detection in multi-turn dialogue is a timely concern as conversational agents are deployed more widely, and the temporal-graph framing is a plausible way to capture cross-turn contradictions. It may interest listeners who care about LLM evaluation and post-hoc safety tooling.",
      "why_not_higher": "This appears to be a narrow dialogue-classification paper with only slight gains over prior methods, not a result that changes mainstream model building, serving, or evaluation practice. The memory angle is essentially absent despite the conversational context language.",
      "downgrade_reasons": [
        "narrow task paper",
        "no direct memory/storage contribution",
        "limited systems relevance",
        "slight improvement over baselines",
        "unclear realism of workloads/models"
      ],
      "what_would_raise_priority": "Stronger evidence on widely used LLM chat benchmarks or production-style conversational logs, plus clearer transfer beyond this specific hallucination-detection setup, would raise it.",
      "one_sentence_episode_hook": "Can a temporal graph over conversation turns catch hallucinations that sentence-by-sentence detectors miss, or is this just another small benchmark bump?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03201",
      "title": "Recursive querying of neural networks via weighted structures",
      "abstract": "Expressive querying of machine learning models - viewed as a form of intentional data - enables their verification and interpretation using declarative languages, thereby making learned representations of data more accessible. Motivated by the querying of feedforward neural networks, we investigate logics for weighted structures. In the absence of a bound on neural network depth, such logics must incorporate recursion; thereto we revisit the functional fixpoint mechanism proposed by Gr\u00e4del and Gurevich. We adopt it in a Datalog-like syntax; we extend normal forms for fixpoint logics to weighted structures; and show an equivalent \"loose\" fixpoint mechanism that allows values of inductively defined weight functions to be overwritten. We propose a \"scalar\" restriction of functional fixpoint logic, of polynomial-time data complexity, and show it can express all PTIME model-agnostic queries over reduced networks with polynomially bounded weights. In contrast, we show that very simple model-agnostic queries are already NP-complete. Finally, we consider transformations of weighted structures by iterated transductions.",
      "authors": [
        "Martin Grohe",
        "Christoph Standke",
        "Juno Steegmans",
        "Jan Van den Bussche"
      ],
      "published_at": "2026-01-06T17:30:44",
      "categories": [
        "cs.LO",
        "cs.AI",
        "cs.DB"
      ],
      "url": "http://arxiv.org/abs/2601.03201v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.4297959804534912,
      "sim_memory": 0.3220950961112976,
      "sim_negative": 0.22655397653579712,
      "broad_relevance": 0.4297959804534912,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.5978538990020752,
      "evidence_score": 0.34931471805599457,
      "direct_memory_relevance": 0.3220950961112976,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7499999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3025122768876577,
      "memory_score": 0.0009462950001086212,
      "quality_score": 0.6227258872223977,
      "bridge_score": 0.0009462950001086212,
      "max_axis_score": 0.3025122768876577,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Theory"
      ],
      "status": "Deferred this cycle",
      "why_now": "It tackles an interesting foundational question: how to declaratively query and reason about neural networks of unbounded depth using recursive weighted logics. That could matter for verification and interpretability communities as AI systems become harder to inspect.",
      "why_not_higher": "The paper is primarily logic/database theory, with little direct connection to how modern large-scale AI systems are built or served. It also does not appear to validate claims on realistic LLMs, multimodal models, or hardware-backed workloads that would make it timely for a broad AI audience.",
      "downgrade_reasons": [
        "theory-heavy with weak near-term practitioner relevance",
        "no direct memory/storage systems contribution",
        "limited evidence on realistic modern models",
        "querying/verification framing is niche for broad audience",
        "unclear impact on current transformer training or inference stacks"
      ],
      "what_would_raise_priority": "A follow-up showing concrete querying or verification wins on real transformer models, with practical tooling and measurable implications for evaluation or serving, would raise it substantially.",
      "one_sentence_episode_hook": "Can we query a neural network like a database\u2014and what does recursion buy us when the model is too deep to reason about with fixed-depth logic?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03390",
      "title": "Revisiting Speculative Leaderless Protocols for Low-Latency BFT Replication",
      "abstract": "As Byzantine Fault Tolerant (BFT) protocols begin to be used in permissioned blockchains for user-facing applications such as payments, it is crucial that they provide low latency. In pursuit of low latency, some recently proposed BFT consensus protocols employ a leaderless optimistic fast path, in which clients broadcast their requests directly to replicas without first serializing requests at a leader, resulting in an end-to-end commit latency of 2 message delays ($2\u0394$) during fault-free, synchronous periods. However, such a fast path only works if there is no contention: concurrent contending requests can cause replicas to diverge if they receive conflicting requests in different orders, triggering costly recovery procedures.   In this work, we present Aspen, a leaderless BFT protocol that achieves a near-optimal latency of $2\u0394+ \\varepsilon$, where $\\varepsilon$ indicates a short waiting delay. Aspen removes the no-contention condition by utilizing a best-effort sequencing layer based on loosely synchronized clocks and network delay estimates. Aspen requires $n = 3f + 2p + 1$ replicas to cope with up to $f$ Byzantine nodes. The $2p$ extra nodes allow Aspen's fast path to proceed even if up to $p$ replicas diverge due to unpredictable network delays. When its optimistic conditions do not hold, Aspen falls back to PBFT-style protocol, guaranteeing safety and liveness under partial synchrony. In experiments with wide-area distributed replicas, Aspen commits requests in less than 75 ms, a 1.2 to 3.3$\\times$ improvement compared to previous protocols, while supporting 19,000 requests per second.",
      "authors": [
        "Daniel Qian",
        "Xiyu Hao",
        "Jinkun Geng",
        "Yuncheng Yao",
        "Aurojit Panda",
        "Jinyang Li",
        "Anirudh Sivaraman"
      ],
      "published_at": "2026-01-06T19:56:15",
      "categories": [
        "cs.DC"
      ],
      "url": "http://arxiv.org/abs/2601.03390v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "systems",
      "domain_bucket": "other",
      "paper_type": "application",
      "narrow_domain_flag": false,
      "sim_public": 0.30004173517227173,
      "sim_memory": 0.3478519022464752,
      "sim_negative": 0.11378025263547897,
      "broad_relevance": 0.30004173517227173,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.5231203436851501,
      "evidence_score": 0.38,
      "direct_memory_relevance": 0.3478519022464752,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.7,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.52,
      "clarity": 0.7,
      "reproducibility": 0.3,
      "public_interest_score": 0.24348057210445406,
      "memory_score": 0.07035557067394252,
      "quality_score": 0.455,
      "bridge_score": 0.07035557067394252,
      "max_axis_score": 0.24348057210445406,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Systems",
        "Application"
      ],
      "status": "Deferred this cycle",
      "why_now": "Low-latency BFT matters for permissioned blockchains and distributed services that want fast commits under adversarial faults, and this paper revisits an active design point around speculative leaderless replication. The WAN evaluation and fallback-to-PBFT design make it timely within distributed systems.",
      "why_not_higher": "This is not really an AI systems paper, and its connection to memory/storage first-class AI is absent beyond generic systems overlap. The main contribution is in BFT replication latency rather than changing how modern AI models are trained, served, or evaluated.",
      "downgrade_reasons": [
        "distributed consensus rather than AI-specific systems",
        "no direct KV cache, offload, bandwidth, paging, or storage mechanism",
        "limited relevance to model training or inference stacks",
        "application domain is permissioned blockchain replication"
      ],
      "what_would_raise_priority": "Show a concrete path where the protocol materially improves distributed AI serving, parameter/state replication, or memory/data-movement behavior in real model-serving deployments.",
      "one_sentence_episode_hook": "Can a leaderless BFT protocol really get near-2\u0394 commits without falling apart under contention\u2014and does that matter for any real AI infrastructure?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ],
  "out_of_scope": [
    {
      "arxiv_id": "2601.03085",
      "title": "Real-Time Adaptive Anomaly Detection in Industrial IoT Environments",
      "abstract": "To ensure reliability and service availability, next-generation networks are expected to rely on automated anomaly detection systems powered by advanced machine learning methods with the capability of handling multi-dimensional data. Such multi-dimensional, heterogeneous data occurs mostly in today's industrial Internet of Things (IIoT), where real-time detection of anomalies is critical to prevent impending failures and resolve them in a timely manner. However, existing anomaly detection methods often fall short of effectively coping with the complexity and dynamism of multi-dimensional data streams in IIoT. In this paper, we propose an adaptive method for detecting anomalies in IIoT streaming data utilizing a multi-source prediction model and concept drift adaptation. The proposed anomaly detection algorithm merges a prediction model into a novel drift adaptation method resulting in accurate and efficient anomaly detection that exhibits improved scalability. Our trace-driven evaluations indicate that the proposed method outperforms the state-of-the-art anomaly detection methods by achieving up to an 89.71% accuracy (in terms of Area under the Curve (AUC)) while meeting the given efficiency and scalability requirements.",
      "authors": [
        "Mahsa Raeiszadeh",
        "Amin Ebrahimzadeh",
        "Roch H. Glitho",
        "Johan Eker",
        "Raquel A. F. Mini"
      ],
      "published_at": "2026-01-06T15:16:45",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03085v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 18,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": true,
      "sim_public": 0.2699599862098694,
      "sim_memory": 0.17710626125335693,
      "sim_negative": 0.4651891589164734,
      "broad_relevance": 0.2699599862098694,
      "momentum": 0.2737531993577084,
      "teachability": 0.85,
      "novelty_score": 0.5759907364845276,
      "evidence_score": 0.634443897916644,
      "direct_memory_relevance": 0.17710626125335693,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.38,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3563038308946783,
      "memory_score": 0.011465146126004355,
      "quality_score": 0.6377775591666576,
      "bridge_score": 0.011465146126004355,
      "max_axis_score": 0.3563038308946783,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [
        "compound_window:180d"
      ],
      "time_window": "180d",
      "compound_window_boost": 0.045,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Out of scope",
      "why_now": "Reliable real-time anomaly detection for industrial IoT is operationally important, especially as automated network management becomes more common. The paper also touches on online adaptation and concept drift, which are timely themes in streaming ML.",
      "why_not_higher": "This is a domain-specific anomaly-detection paper rather than something that changes mainstream AI practice or the memory/storage stack of modern models. The abstract suggests incremental empirical improvement, but not a broadly transferable idea likely to matter to a wide AI audience in the next 30-90 days.",
      "downgrade_reasons": [
        "narrow IIoT application focus",
        "no direct memory/storage mechanism",
        "limited relevance to modern foundation-model build/serve workflows",
        "benchmark-style contribution from abstract",
        "unclear generalization beyond industrial streaming anomaly detection"
      ],
      "what_would_raise_priority": "Show that the drift-adaptation method generalizes across multiple streaming AI settings and materially changes deployment efficiency or robustness in broader production ML systems.",
      "one_sentence_episode_hook": "Can adaptive concept-drift detection make industrial IoT anomaly monitoring truly real-time, or is this mostly a niche benchmark win?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03232",
      "title": "Multi-RADS Synthetic Radiology Report Dataset and Head-to-Head Benchmarking of 41 Open-Weight and Proprietary Language Models",
      "abstract": "Background: Reporting and Data Systems (RADS) standardize radiology risk communication but automated RADS assignment from narrative reports is challenging because of guideline complexity, output-format constraints, and limited benchmarking across RADS frameworks and model sizes. Purpose: To create RXL-RADSet, a radiologist-verified synthetic multi-RADS benchmark, and compare validity and accuracy of open-weight small language models (SLMs) with a proprietary model for RADS assignment. Materials and Methods: RXL-RADSet contains 1,600 synthetic radiology reports across 10 RADS (BI-RADS, CAD-RADS, GB-RADS, LI-RADS, Lung-RADS, NI-RADS, O-RADS, PI-RADS, TI-RADS, VI-RADS) and multiple modalities. Reports were generated by LLMs using scenario plans and simulated radiologist styles and underwent two-stage radiologist verification. We evaluated 41 quantized SLMs (12 families, 0.135-32B parameters) and GPT-5.2 under a fixed guided prompt. Primary endpoints were validity and accuracy; a secondary analysis compared guided versus zero-shot prompting. Results: Under guided prompting GPT-5.2 achieved 99.8% validity and 81.1% accuracy (1,600 predictions). Pooled SLMs (65,600 predictions) achieved 96.8% validity and 61.1% accuracy; top SLMs in the 20-32B range reached ~99% validity and mid-to-high 70% accuracy. Performance scaled with model size (inflection between <1B and >=10B) and declined with RADS complexity primarily due to classification difficulty rather than invalid outputs. Guided prompting improved validity (99.2% vs 96.7%) and accuracy (78.5% vs 69.6%) compared with zero-shot. Conclusion: RXL-RADSet provides a radiologist-verified multi-RADS benchmark; large SLMs (20-32B) can approach proprietary-model performance under guided prompting, but gaps remain for higher-complexity schemes.",
      "authors": [
        "Kartik Bose",
        "Abhinandan Kumar",
        "Raghuraman Soundararajan",
        "Priya Mudgil",
        "Samonee Ralmilay",
        "Niharika Dutta",
        "Manphool Singhal",
        "Arun Kumar",
        "Saugata Sen",
        "Anurima Patra"
      ],
      "published_at": "2026-01-06T18:18:44",
      "categories": [
        "cs.CL",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03232v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.3113279342651367,
      "sim_memory": 0.2988414168357849,
      "sim_negative": 0.333227276802063,
      "broad_relevance": 0.3113279342651367,
      "momentum": 0.017490950964499406,
      "teachability": 0.75,
      "novelty_score": 0.70535808801651,
      "evidence_score": 0.5293147180559945,
      "direct_memory_relevance": 0.2988414168357849,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6599999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3430974913833166,
      "memory_score": 0.0009701912174548133,
      "quality_score": 0.6027258872223977,
      "bridge_score": 0.0009701912174548133,
      "max_axis_score": 0.3430974913833166,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Out of scope",
      "why_now": "It is a timely, reasonably well-executed benchmark of open versus proprietary LLMs on structured radiology report classification, with a useful finding that guided prompting materially improves validity and accuracy. The dataset spans multiple RADS schemes and many model sizes, which makes it a solid domain benchmark.",
      "why_not_higher": "The contribution is mostly a synthetic medical benchmark plus head-to-head model ranking, not a result that broadly changes mainstream AI practice. Its memory/storage relevance is essentially absent, and the transfer beyond this narrow clinical labeling task is limited.",
      "downgrade_reasons": [
        "narrow domain application",
        "benchmark-heavy without broader systems takeaway",
        "no direct memory/storage mechanism",
        "limited transferability outside radiology structured classification",
        "synthetic dataset reduces real-world operational relevance"
      ],
      "what_would_raise_priority": "Show that the benchmark reveals broadly transferable prompting, constrained decoding, or small-model deployment lessons that generalize to high-stakes structured generation beyond radiology.",
      "one_sentence_episode_hook": "Can 20B\u201332B open models nearly match a frontier proprietary model on radiology risk scoring\u2014and does prompt scaffolding matter more than sheer scale for getting the format right?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.04257",
      "title": "Cross-Language Speaker Attribute Prediction Using MIL and RL",
      "abstract": "We study multilingual speaker attribute prediction under linguistic variation, domain mismatch, and data imbalance across languages. We propose RLMIL-DAT, a multilingual extension of the reinforced multiple instance learning framework that combines reinforcement learning based instance selection with domain adversarial training to encourage language invariant utterance representations. We evaluate the approach on a five language Twitter corpus in a few shot setting and on a VoxCeleb2 derived corpus covering forty languages in a zero shot setting for gender and age prediction. Across a wide range of model configurations and multiple random seeds, RLMIL-DAT consistently improves Macro F1 compared to standard multiple instance learning and the original reinforced multiple instance learning framework. The largest gains are observed for gender prediction, while age prediction remains more challenging and shows smaller but positive improvements. Ablation experiments indicate that domain adversarial training is the primary contributor to the performance gains, enabling effective transfer from high resource English to lower resource languages by discouraging language specific cues in the shared encoder. In the zero shot setting on the smaller VoxCeleb2 subset, improvements are generally positive but less consistent, reflecting limited statistical power and the difficulty of generalizing to many unseen languages. Overall, the results demonstrate that combining instance selection with adversarial domain adaptation is an effective and robust strategy for cross lingual speaker attribute prediction.",
      "authors": [
        "Sunny Shu",
        "Seyed Sahand Mohammadi Ziabari",
        "Ali Mohammed Mansoor Alsahag"
      ],
      "published_at": "2026-01-06T23:07:26",
      "categories": [
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.04257v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.41569265723228455,
      "sim_memory": 0.3652263879776001,
      "sim_negative": 0.18383526802062988,
      "broad_relevance": 0.41569265723228455,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.5844208598136902,
      "evidence_score": 0.35000000000000003,
      "direct_memory_relevance": 0.3652263879776001,
      "systems_leverage": 0.1,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.7,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3423709261417389,
      "memory_score": 0.005567916393280026,
      "quality_score": 0.595,
      "bridge_score": 0.005567916393280026,
      "max_axis_score": 0.3423709261417389,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Out of scope",
      "why_now": "Cross-lingual robustness and few-shot transfer remain relevant AI themes, and the paper combines RL-based instance selection with domain adversarial training for multilingual speaker attribute prediction. It is timely as a practical attempt to transfer from high-resource English to lower-resource languages.",
      "why_not_higher": "This is a narrow speech-attribute application paper rather than something that changes mainstream AI system design or evaluation. Its memory/storage relevance is effectively absent, and the empirical evidence is limited to task-specific datasets with modest generalization claims.",
      "downgrade_reasons": [
        "No direct memory/storage mechanism",
        "Narrow task-specific application",
        "Limited impact on how modern AI systems are built or served",
        "Evidence comes from specialized datasets rather than broadly realistic frontier workloads",
        "Episode would likely center on benchmark deltas rather than a broadly reusable idea"
      ],
      "what_would_raise_priority": "Stronger evidence that the method transfers broadly across multimodal or foundation-model settings, or a concrete systems/memory implication, would raise priority.",
      "one_sentence_episode_hook": "Can adversarial training plus RL-style utterance selection make speaker prediction actually transfer across dozens of languages, or is this mostly a niche benchmark win?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03098",
      "title": "From Muscle to Text with MyoText: sEMG to Text via Finger Classification and Transformer-Based Decoding",
      "abstract": "Surface electromyography (sEMG) provides a direct neural interface for decoding muscle activity and offers a promising foundation for keyboard-free text input in wearable and mixed-reality systems. Previous sEMG-to-text studies mainly focused on recognizing letters directly from sEMG signals, forming an important first step toward translating muscle activity into text. Building on this foundation, we present MyoText, a hierarchical framework that decodes sEMG signals to text through physiologically grounded intermediate stages. MyoText first classifies finger activations from multichannel sEMG using a CNN-BiLSTM-Attention model, applies ergonomic typing priors to infer letters, and reconstructs full sentences with a fine-tuned T5 transformer. This modular design mirrors the natural hierarchy of typing, linking muscle intent to language output and reducing the search space for decoding. Evaluated on 30 users from the emg2qwerty dataset, MyoText outperforms baselines by achieving 85.4% finger-classification accuracy, 5.4% character error rate (CER), and 6.5% word error rate (WER). Beyond accuracy gains, this methodology establishes a principled pathway from neuromuscular signals to text, providing a blueprint for virtual and augmented-reality typing interfaces that operate entirely without physical keyboards. By integrating ergonomic structure with transformer-based linguistic reasoning, MyoText advances the feasibility of seamless, wearable neural input for future ubiquitous computing environments.",
      "authors": [
        "Meghna Roy Chowdhury",
        "Shreyas Sen",
        "Yi Ding"
      ],
      "published_at": "2026-01-06T15:30:15",
      "categories": [
        "cs.LG",
        "cs.NE"
      ],
      "url": "http://arxiv.org/abs/2601.03098v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.3751242160797119,
      "sim_memory": 0.33207595348358154,
      "sim_negative": 0.25948554277420044,
      "broad_relevance": 0.3751242160797119,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6275269389152527,
      "evidence_score": 0.43000000000000005,
      "direct_memory_relevance": 0.33207595348358154,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3366663056612015,
      "memory_score": 0.00262278604507446,
      "quality_score": 0.575,
      "bridge_score": 0.00262278604507446,
      "max_axis_score": 0.3366663056612015,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Out of scope",
      "why_now": "It shows a neat hierarchical pipeline from biosignals to text, combining physiological priors with a transformer decoder in a way that is timely for wearable and XR input interfaces. The modular decomposition is more interesting than a pure end-to-end benchmark bump.",
      "why_not_higher": "This is still a narrow sEMG typing application, not a paper that is likely to change mainstream AI practice in the next 30-90 days. Its transformer use is downstream language reconstruction rather than a broadly transferable advance in model design or systems.",
      "downgrade_reasons": [
        "narrow application domain",
        "no direct memory/storage contribution",
        "limited impact on mainstream AI workflows",
        "transferability beyond sEMG typing is unclear"
      ],
      "what_would_raise_priority": "Show that the hierarchical decoding recipe generalizes across neural or multimodal input interfaces and alters how broader AI systems are built or evaluated.",
      "one_sentence_episode_hook": "Can muscle signals become fluent text if you decode fingers first and let a language model fill in the sentence?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03062",
      "title": "Explainable Fuzzy GNNs for Leak Detection in Water Distribution Networks",
      "abstract": "Timely leak detection in water distribution networks is critical for conserving resources and maintaining operational efficiency. Although Graph Neural Networks (GNNs) excel at capturing spatial-temporal dependencies in sensor data, their black-box nature and the limited work on graph-based explainable models for water networks hinder practical adoption. We propose an explainable GNN framework that integrates mutual information to identify critical network regions and fuzzy logic to provide clear, rule-based explanations for node classification tasks. After benchmarking several GNN architectures, we selected the generalized graph convolution network (GENConv) for its superior performance and developed a fuzzy-enhanced variant that offers intuitive explanations for classified leak locations. Our fuzzy graph neural network (FGENConv) achieved Graph F1 scores of 0.889 for detection and 0.814 for localization, slightly below the crisp GENConv 0.938 and 0.858, respectively. Yet it compensates by providing spatially localized, fuzzy rule-based explanations. By striking the right balance between precision and explainability, the proposed fuzzy network could enable hydraulic engineers to validate predicted leak locations, conserve human resources, and optimize maintenance strategies. The code is available at github.com/pasqualedem/GNNLeakDetection.",
      "authors": [
        "Qusai Khaled",
        "Pasquale De Marinis",
        "Moez Louati",
        "David Ferras",
        "Laura Genga",
        "Uzay Kaymak"
      ],
      "published_at": "2026-01-06T14:45:43",
      "categories": [
        "cs.AI",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03062v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.3916358947753906,
      "sim_memory": 0.152069553732872,
      "sim_negative": 0.32004913687705994,
      "broad_relevance": 0.3916358947753906,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6531769037246704,
      "evidence_score": 0.36000000000000004,
      "direct_memory_relevance": 0.152069553732872,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6799999999999999,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.3254673039913178,
      "memory_score": 0.008620866119861598,
      "quality_score": 0.595,
      "bridge_score": 0.008620866119861598,
      "max_axis_score": 0.3254673039913178,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Out of scope",
      "why_now": "It addresses a real infrastructure problem and adds explainability to graph leak detection, which could matter to practitioners in water networks. The explainable-GNN angle is timely, but mainly within this application niche rather than the broader AI conversation.",
      "why_not_higher": "This is a narrow domain paper with limited relevance to how the wider AI field builds or serves modern models. Its memory/storage connection is essentially absent in the podcast's intended systems sense.",
      "downgrade_reasons": [
        "narrow application focus",
        "no direct memory/storage systems contribution",
        "limited transfer beyond water-network leak detection",
        "mostly benchmark-plus-explainability tradeoff"
      ],
      "what_would_raise_priority": "Show that the fuzzy-explainable graph method generalizes across multiple high-impact graph domains and meaningfully changes model evaluation or deployment practice.",
      "one_sentence_episode_hook": "Can a fuzzy, explainable GNN make leak detection trustworthy enough for real water networks\u2014or is this just a niche benchmark with nicer explanations?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03376",
      "title": "Weather-Aware Transformer for Real-Time Route Optimization in Drone-as-a-Service Operations",
      "abstract": "This paper presents a novel framework to accelerate route prediction in Drone-as-a-Service operations through weather-aware deep learning models. While classical path-planning algorithms, such as A* and Dijkstra, provide optimal solutions, their computational complexity limits real-time applicability in dynamic environments. We address this limitation by training machine learning and deep learning models on synthetic datasets generated from classical algorithm simulations. Our approach incorporates transformer-based and attention-based architectures that utilize weather heuristics to predict optimal next-node selections while accounting for meteorological conditions affecting drone operations. The attention mechanisms dynamically weight environmental factors including wind patterns, wind bearing, and temperature to enhance routing decisions under adverse weather conditions. Experimental results demonstrate that our weather-aware models achieve significant computational speedup over traditional algorithms while maintaining route optimization performance, with transformer-based architectures showing superior adaptation to dynamic environmental constraints. The proposed framework enables real-time, weather-responsive route optimization for large-scale DaaS operations, representing a substantial advancement in the efficiency and safety of autonomous drone systems.",
      "authors": [
        "Kamal Mohamed",
        "Lillian Wassim",
        "Ali Hamdi",
        "Khaled Shaban"
      ],
      "published_at": "2026-01-06T19:23:15",
      "categories": [
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.03376v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "pde",
      "paper_type": "empirical",
      "narrow_domain_flag": true,
      "sim_public": 0.47454923391342163,
      "sim_memory": 0.30600279569625854,
      "sim_negative": 0.5202996730804443,
      "broad_relevance": 0.47454923391342163,
      "momentum": 0.0,
      "teachability": 0.75,
      "novelty_score": 0.6510633826255798,
      "evidence_score": 0.32,
      "direct_memory_relevance": 0.30600279569625854,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.25,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.38,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.3178444737195969,
      "memory_score": 0.027621034860610932,
      "quality_score": 0.5,
      "bridge_score": 0.027621034860610932,
      "max_axis_score": 0.3178444737195969,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "90d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0121798038482666,
      "badges": [
        "Application",
        "Inference"
      ],
      "status": "Out of scope",
      "why_now": "It touches a timely theme\u2014using learned policies to replace slower classical optimization under real-time constraints\u2014and packages it with weather-aware attention for autonomous drone routing. That could interest specialists in robotics or DaaS operations.",
      "why_not_higher": "For this show's lenses, it looks like a narrow application paper rather than something that changes mainstream AI practice or memory/storage behavior. The evidence also appears to rely on synthetic data and classical-simulation labels rather than realistic large-scale deployment or systems evaluation.",
      "downgrade_reasons": [
        "narrow task-specific routing application",
        "memory/storage link absent",
        "transformer use appears incidental rather than field-shaping",
        "synthetic-data evaluation limits confidence",
        "unclear impact on modern AI system design"
      ],
      "what_would_raise_priority": "A stronger case would show a transferable architecture or systems result that improves real-world routing under realistic hardware, fleet-scale workloads, and broadly reusable decision-making settings.",
      "one_sentence_episode_hook": "Can a weather-aware transformer learn to beat classical route planning fast enough for real-time drone fleets\u2014or is this just another narrow task dressed in transformer language?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03173",
      "title": "Predicting Time Pressure of Powered Two-Wheeler Riders for Proactive Safety Interventions",
      "abstract": "Time pressure critically influences risky maneuvers and crash proneness among powered two-wheeler riders, yet its prediction remains underexplored in intelligent transportation systems. We present a large-scale dataset of 129,000+ labeled multivariate time-series sequences from 153 rides by 51 participants under No, Low, and High Time Pressure conditions. Each sequence captures 63 features spanning vehicle kinematics, control inputs, behavioral violations, and environmental context. Our empirical analysis shows High Time Pressure induces 48% higher speeds, 36.4% greater speed variability, 58% more risky turns at intersections, 36% more sudden braking, and 50% higher rear brake forces versus No Time Pressure. To benchmark this dataset, we propose MotoTimePressure, a deep learning model combining convolutional preprocessing, dual-stage temporal attention, and Squeeze-and-Excitation feature recalibration, achieving 91.53% accuracy and 98.93% ROC AUC, outperforming eight baselines. Since time pressure cannot be directly measured in real time, we demonstrate its utility in collision prediction and threshold determination. Using MTPS-predicted time pressure as features, improves Informer-based collision risk accuracy from 91.25% to 93.51%, approaching oracle performance (93.72%). Thresholded time pressure states capture rider cognitive stress and enable proactive ITS interventions, including adaptive alerts, haptic feedback, V2I signaling, and speed guidance, supporting safer two-wheeler mobility under the Safe System Approach.",
      "authors": [
        "Sumit S. Shevtekar",
        "Chandresh K. Maurya",
        "Gourab Sil",
        "Subasish Das"
      ],
      "published_at": "2026-01-06T16:52:09",
      "categories": [
        "cs.LG",
        "cs.HC"
      ],
      "url": "http://arxiv.org/abs/2601.03173v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.33928826451301575,
      "sim_memory": 0.25232285261154175,
      "sim_negative": 0.2517510950565338,
      "broad_relevance": 0.33928826451301575,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6552061438560486,
      "evidence_score": 0.44,
      "direct_memory_relevance": 0.25232285261154175,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.45999999999999996,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.31006740093231205,
      "memory_score": 0.0,
      "quality_score": 0.545,
      "bridge_score": 0.0,
      "max_axis_score": 0.31006740093231205,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Out of scope",
      "why_now": "Road-safety sensing and proactive intervention are timely, and the paper offers a sizable labeled rider-behavior dataset with a strong benchmark result on predicting time pressure. It may interest ITS and human-centered ML communities looking at stress-aware assistance.",
      "why_not_higher": "This is a narrow transportation application paper rather than something that changes mainstream AI practice, and its memory/storage relevance is effectively absent. The modeling contribution looks like a task-specific architecture-and-benchmark package with limited transfer to modern foundation-model training, serving, or evaluation.",
      "downgrade_reasons": [
        "narrow domain application",
        "no direct memory or storage mechanism",
        "limited relevance to modern AI systems building/serving",
        "benchmark-centric contribution",
        "task-specific architecture"
      ],
      "what_would_raise_priority": "Show that the time-pressure modeling or intervention framework transfers broadly across multimodal sequential decision systems or introduces a reusable evaluation pattern adopted beyond two-wheeler safety.",
      "one_sentence_episode_hook": "Can an AI system infer when a motorcycle rider is under time pressure early enough to intervene before risky behavior turns into a crash?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.03327",
      "title": "Extreme-value forest fire prediction A study of the Loss Function in an Ordinality Scheme",
      "abstract": "Wildfires are highly imbalanced natural hazards in both space and severity, making the prediction of extreme events particularly challenging. In this work, we introduce the first ordinal classification framework for forecasting wildfire severity levels directly aligned with operational decision-making in France. Our study investigates the influence of loss-function design on the ability of neural models to predict rare yet critical high-severity fire occurrences. We compare standard cross-entropy with several ordinal-aware objectives, including the proposed probabilistic TDeGPD loss derived from a truncated discrete exponentiated Generalized Pareto Distribution. Through extensive benchmarking over multiple architectures and real operational data, we show that ordinal supervision substantially improves model performance over conventional approaches. In particular, the Weighted Kappa Loss (WKLoss) achieves the best overall results, with more than +0.1 IoU (Intersection Over Union) gain on the most extreme severity classes while maintaining competitive calibration quality. However, performance remains limited for the rarest events due to their extremely low representation in the dataset. These findings highlight the importance of integrating both severity ordering, data imbalance considerations, and seasonality risk into wildfire forecasting systems. Future work will focus on incorporating seasonal dynamics and uncertainty information into training to further improve the reliability of extreme-event prediction.",
      "authors": [
        "Nicolas Caron",
        "Christophe Guyeux",
        "Hassan Noura",
        "Benjamin Aynes"
      ],
      "published_at": "2026-01-06T15:46:56",
      "categories": [
        "cs.LG",
        "cs.AI"
      ],
      "url": "http://arxiv.org/abs/2601.03327v2",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 1,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "other",
      "paper_type": "benchmark",
      "narrow_domain_flag": false,
      "sim_public": 0.3068709969520569,
      "sim_memory": 0.2279648780822754,
      "sim_negative": 0.32376688718795776,
      "broad_relevance": 0.3068709969520569,
      "momentum": 0.017490950964499406,
      "teachability": 0.85,
      "novelty_score": 0.6007909774780273,
      "evidence_score": 0.5093147180559946,
      "direct_memory_relevance": 0.2279648780822754,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.6599999999999999,
      "clarity": 0.7,
      "reproducibility": 0.5,
      "public_interest_score": 0.30607534360862027,
      "memory_score": 0.0,
      "quality_score": 0.6027258872223977,
      "bridge_score": 0.0,
      "max_axis_score": 0.30607534360862027,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application",
        "Training"
      ],
      "status": "Out of scope",
      "why_now": "It addresses a real and socially important forecasting problem and offers a useful comparison of ordinal-aware losses for rare extreme-event prediction on operational wildfire data. The loss-function focus is timely for practitioners working on imbalanced ordinal risk tasks.",
      "why_not_higher": "This is primarily a narrow domain forecasting paper rather than one that changes mainstream AI practice or connects to memory/storage systems. The main contribution appears to be benchmarked loss-design choices for a specific operational task, with limited evidence of broad transfer beyond wildfire severity prediction.",
      "downgrade_reasons": [
        "narrow application paper",
        "no direct memory/storage relevance",
        "limited impact on modern AI system design",
        "benchmark-heavy contribution",
        "unclear transfer to foundation-model workflows"
      ],
      "what_would_raise_priority": "Show that the ordinal-loss framework consistently improves rare-event prediction across multiple high-impact domains or modern large-model settings, with clear implications for general training practice.",
      "one_sentence_episode_hook": "Can a better loss function make AI noticeably better at forecasting the rarest, most dangerous wildfires\u2014or is this gain mostly confined to one specialized benchmark?",
      "source": "digest",
      "added": "",
      "issue_number": null
    },
    {
      "arxiv_id": "2601.02965",
      "title": "Low-Resource Heuristics for Bahnaric Optical Character Recognition Improvement",
      "abstract": "Bahnar, a minority language spoken across Vietnam, Cambodia, and Laos, faces significant preservation challenges due to limited research and data availability. This study addresses the critical need for accurate digitization of Bahnar language documents through optical character recognition (OCR) technology. Digitizing scanned paper documents poses significant challenges, as degraded image quality from broken or blurred areas introduces considerable OCR errors that compromise information retrieval systems. We propose a comprehensive approach combining advanced table and non-table detection techniques with probability-based post-processing heuristics to enhance recognition accuracy. Our method first applies detection algorithms to improve input data quality, then employs probabilistic error correction on OCR output. Experimental results indicate a substantial improvement, with recognition accuracy increasing from 72.86% to 79.26%. This work contributes valuable resources for Bahnar language preservation and provides a framework applicable to other minority language digitization efforts.",
      "authors": [
        "Phat Tran",
        "Phuoc Pham",
        "Hung Trinh",
        "Tho Quan"
      ],
      "published_at": "2026-01-06T12:22:03",
      "categories": [
        "cs.CL",
        "cs.CV",
        "cs.LG"
      ],
      "url": "http://arxiv.org/abs/2601.02965v1",
      "code_url": "",
      "github_submission_flag": false,
      "hf_trending_flag": false,
      "citation_count": 0,
      "influential_citation_count": 0,
      "scope_bucket": "foundation",
      "domain_bucket": "llm",
      "paper_type": "empirical",
      "narrow_domain_flag": false,
      "sim_public": 0.294531911611557,
      "sim_memory": 0.2657720446586609,
      "sim_negative": 0.26265424489974976,
      "broad_relevance": 0.294531911611557,
      "momentum": 0.0,
      "teachability": 0.85,
      "novelty_score": 0.6129536628723145,
      "evidence_score": 0.34,
      "direct_memory_relevance": 0.2657720446586609,
      "systems_leverage": 0.25,
      "deployment_proximity": 0.2,
      "memory_adjacent_future_value": 0.0,
      "bandwidth_capacity": 0.0,
      "transferability_score": 0.72,
      "clarity": 0.7999999999999999,
      "reproducibility": 0.5,
      "public_interest_score": 0.27030262291431423,
      "memory_score": 0.0027316133975982537,
      "quality_score": 0.595,
      "bridge_score": 0.0027316133975982537,
      "max_axis_score": 0.27030262291431423,
      "influencer_boost": 0.0,
      "influencer_matches": [],
      "pwc_trending_flag": false,
      "social_score": 0.0,
      "scoring_sources": [],
      "time_window": "180d",
      "compound_window_boost": 0.0,
      "first_seen_date": "",
      "fatigue_penalty": 0.0,
      "negative_profile_penalty": 0.0,
      "badges": [
        "Application"
      ],
      "status": "Out of scope",
      "why_now": "It addresses a real preservation problem for a low-resource language and shows a measurable OCR accuracy gain from pragmatic detection and post-processing heuristics. That could matter to niche document AI and language preservation communities.",
      "why_not_higher": "This is a narrow OCR improvement paper rather than something that changes mainstream AI practice, and the memory/storage connection is absent. The method appears task-specific and heuristic-driven, with limited evidence of broad transfer beyond similar minority-language document pipelines.",
      "downgrade_reasons": [
        "narrow application focus",
        "no direct memory/storage relevance",
        "limited impact on modern AI system design",
        "heuristic post-processing rather than general method",
        "modest empirical scope"
      ],
      "what_would_raise_priority": "Show strong cross-language transfer on multiple OCR benchmarks and connect the approach to generally reusable document AI evaluation or deployment practices.",
      "one_sentence_episode_hook": "Can low-resource heuristics meaningfully rescue OCR for endangered languages, or is this a useful but highly local document-processing tweak?",
      "source": "digest",
      "added": "",
      "issue_number": null
    }
  ]
}