YouRA: A Persistent-State Architecture for Evidence-Traceable Autonomous Research Agents

AACL-IJCNLP 2026 (Main) · arXiv:2610.01097

Abstract

End-to-end research agents can now produce complete scientific papers, but manuscript claims can diverge from executed experiments. YouRA (Your Research Agent) addresses this reliability gap by making research state, execution evidence, and trajectory-level failure history explicit, persistent, and verifiable across the full research lifecycle.

YouRA combines three architectural components:

Component	Description
Verification State Architecture (VSA)	Tracks hypotheses, gates, evidence pointers, checkpoints, tasks, reflective memory, and sequential memory as persistent state.
Independent Controller	Reads VSA summaries, per-phase outputs, and reflective memory to control lifecycle continuation, recovery planning, and independent debate/review.
Stateful Reflection	Records failures as structured lessons and routes recovery through bounded repair, hypothesis redesign, or problem-scoping reset.

Given only a topic-level description, YouRA expands it into a research question, hypothesis, experiment plan, implementation, validation evidence, and final manuscript. For benchmark evaluation, runs are held to fully autonomous mode; for practical use, the same infrastructure exposes stage-specific CLI commands and resumable hook-driven orchestration.

Key Results

The paper evaluates YouRA on the MLR-Bench predefined ten-task end-to-end subset.

End-to-End Scores

MLR-Bench scores use a 1-10 rubric where higher is better. Each cell reports mean ± task standard deviation over four judges.

System	Backbone LLM	Clarity	Novelty	Soundness	Significance	Overall
MLR-Agent	Sonnet 4.5	7.42 ± 0.24	5.12 ± 0.74	2.48 ± 0.61	3.52 ± 0.68	3.10 ± 0.64
MLR-Agent	Opus 4.5	7.30 ± 0.35	5.25 ± 0.31	3.05 ± 0.96	3.80 ± 0.71	3.62 ± 0.80
MLR-Agent	Sonnet 4.6	7.78 ± 0.22	5.95 ± 0.44	4.05 ± 1.19	4.72 ± 1.14	4.42 ± 1.11
AI Scientist V2	Sonnet 4.5	6.72 ± 0.43	4.85 ± 0.60	3.45 ± 1.18	3.92 ± 0.85	3.62 ± 0.84
AI Scientist V2	Opus 4.5	6.38 ± 1.25	5.22 ± 0.86	4.18 ± 1.31	4.47 ± 1.03	4.28 ± 1.16
AI Scientist V2	Sonnet 4.6	6.45 ± 0.65	5.75 ± 0.47	3.32 ± 1.01	4.22 ± 0.95	3.88 ± 0.88
YouRA	Sonnet 4.5	7.50 ± 0.49	5.23 ± 0.92	3.98 ± 1.74	4.30 ± 1.28	4.20 ± 1.45
YouRA	Opus 4.5	7.73 ± 0.48	5.33 ± 1.03	4.58 ± 1.86	4.33 ± 1.42	4.45 ± 1.51
YouRA	Sonnet 4.6	7.72 ± 0.32	5.70 ± 0.57	4.92 ± 1.09	5.05 ± 1.00	5.05 ± 0.92

Pairwise Preferences

Each row summarizes 40 order-collapsed judge-task verdicts from the YouRA perspective (10 tasks x 4 judges). Each verdict is evaluated in both presentation orders; order reversals are counted as ties.

Comparison	Backbone	Win	Tie	Lose
YouRA vs MLR-Agent	Sonnet 4.5	25	14	1
YouRA vs MLR-Agent	Opus 4.5	26	5	9
YouRA vs MLR-Agent	Sonnet 4.6	21	14	5
YouRA vs AI Scientist V2	Sonnet 4.5	18	14	8
YouRA vs AI Scientist V2	Opus 4.5	17	12	11
YouRA vs AI Scientist V2	Sonnet 4.6	20	12	8

Architecture Overview

YouRA's lifecycle proceeds left to right under the VSA. The independent controller consumes persistent state and reflective memory to decide whether to approve termination, emit auto-resume prompts, or select recovery routing. Sub-hypotheses move through an iterative experiment-design, implementation-planning, and code-validation loop. Failed gates are routed through repair, redesign, or reset, while final evidence passes through manuscript drafting, adversarial fact review, and refinement.

Installation

Prerequisites

• A Linux environment for the unattended launchers (run_phase*.py stream the Claude CLI through POSIX select() on pipes and use /proc for hang detection). On Windows, run the pipeline under WSL.
• Python 3.10+
• Claude Code CLI at exactly ~/.local/bin/claude. Every launcher hard-codes this path (CLAUDE_CLI in YOURA/.claude/hooks/run_phase*.py); if the CLI is installed elsewhere, symlink it there (see Claude Hook Setup).
• Claude Code CLI must be logged in and backed by an active Claude subscription or API-backed account before running YouRA.
• Codex CLI must be installed, logged in, and backed by an active subscription or API key before running Codex-backed evaluation scripts.
• OPENROUTER_API_KEY in .env for the GPT-5.2-based auto-responder/controller
• A TeX toolchain providing pdflatex, xelatex, and bibtex. Phase 6.5.1 compiles the Overleaf project with pdflatex + bibtex; the --enable-refine pass compiles the refined manuscript with xelatex + bibtex. YOURA/setup_tex.py installs TinyTeX plus the required packages (see TeX Toolchain Setup). Without a TeX toolchain, Phase 6.5.1 still writes the .tex project (its verifier treats output.pdf as optional), and the refine pass writes 06_paper_refinement.md and the .tex sources, logs that xelatex/bibtex are missing, and completes without a PDF.
• conda (miniforge3/miniconda): Phase 4 creates a conda environment for each experiment and stops if conda is not found.
• MCP services: Serena (strongly recommended) and Archon (recommended). The phase workflows declare both as required tool servers and the prompts instruct the model to use them for Reflective Memory, task tracking, and knowledge-base search; the Python launchers do not check for them, and runs complete without them with reduced tool grounding (the *_no_mcp ablation lanes ran with no MCP servers). Other optional MCP services can be found in the Smithery server directory.

Setup

Install in this order; each step is detailed below.

1. Python virtual environment, then pip install -e . at the repository root (installs the mlrbench-youra evaluation/analysis package).
2. .env at the repository root with OPENROUTER_API_KEY.
3. cd YOURA && python install_hooks.py --install-deps — installs the hook runtime dependencies from YOURA/requirements.txt into the same virtual environment and writes .claude/settings.local.json (Claude Hook Setup).
4. cd YOURA && python setup_tex.py, then add the TinyTeX bin directory to PATH (TeX Toolchain Setup). Standard library only, so it does not depend on the virtual environment; needed only for PDF output.
5. conda on PATH for the Phase 4 experiment environments.

# Clone repository
git clone https://github.com/PrayPrey/Your-Research-Agent.git
cd Your-Research-Agent

# Create virtual environment
python -m venv .venv
source .venv/bin/activate        # Linux/Mac
# or: .venv\Scripts\activate     # Windows

# Install root evaluation package
pip install -e .

# Configure environment
cp .env.example .env
# then edit .env and fill in OPENROUTER_API_KEY
# OPENAI_API_KEY is optional; needed only for direct OpenAI/Codex API calls

This repository has two Python setup layers plus one non-Python step:

• pyproject.toml installs the root evaluation/analysis package under src/mlrbench.
• YOURA/install_hooks.py installs the small Claude Code hook runtime dependencies (requests, PyYAML, python-dotenv, openai, pytest) and writes repository-specific absolute paths into YOURA/.claude/settings.local.json.
• YOURA/setup_tex.py installs the TeX toolchain used for PDF generation (see TeX Toolchain Setup).

Run both Python steps in the same virtual environment. Do not manually edit YOURA/.claude/settings.local.json; regenerate it with install_hooks.py after moving or recloning the repository. The hooks read .env from YOURA/ first and then from the repository root, so a single .env at the root is enough (YOURA/.env.example is a copy of the root template).

ANTHROPIC_API_KEY is not required for the default YouRA workflow when Claude Code CLI is already logged in. It is only needed if you run the MLR-Bench evaluators with direct Anthropic API models instead of OpenRouter/Claude Code CLI. The default model choices work without additional environment variables; override them only when needed via CLI flags such as --judge-model, --model, or the optional LLM_JUDGE_MODEL, CLAUDE_MODEL, and CODEX_MODEL variables.

Claude Hook Setup

Run the hook installer from the YOURA/ directory. It writes absolute Python paths into YOURA/.claude/settings.local.json, so rerun it after moving the repository.

cd YOURA
python install_hooks.py --install-deps

--claude-bin only tells the installer which binary to version-check; the path is not written to any config. The launchers themselves always call ~/.local/bin/claude, so if the CLI is installed elsewhere, put a symlink at that path before running the pipeline:

mkdir -p ~/.local/bin
ln -s "$(command -v claude)" ~/.local/bin/claude
python install_hooks.py --install-deps

On Windows, the unattended launchers do not run natively (see Prerequisites); use WSL and the Linux instructions above. The installer alone can still be run on Windows to check a claude.EXE install by passing its absolute path with --claude-bin "C:\Users\<you>\.local\bin\claude.EXE".

To skip the Claude CLI check entirely (e.g., when only configuring on a machine where Claude Code will be installed later), add --skip-claude-check.

TeX Toolchain Setup

Phase 6.5.1 compiles the Overleaf project with pdflatex + bibtex, and the --enable-refine pass compiles the refined manuscript with xelatex + bibtex. If an existing TeX Live installation already provides these binaries, nothing else is needed. Otherwise install TinyTeX plus the required packages with the bundled script (each agent folder ships an identical copy):

cd YOURA
python setup_tex.py                                    # TinyTeX under ~/.TinyTeX, then tlmgr packages
export PATH="$HOME/.TinyTeX/bin/x86_64-linux:$PATH"   # add to ~/.bashrc; aarch64-linux / universal-darwin on other platforms
pdflatex --version && xelatex --version && bibtex --version

The script prints a verification block at the end; every entry must resolve to a path (the script exits 0 even when something is NOT FOUND, so check the block). Useful options:

• --dry-run prints the commands without running them.
• --skip-install-tinytex reuses an existing tlmgr and installs only the packages.
• --packages ... overrides the package list. The default list includes CJK support (kotex, cjk, xecjk) and the large collection-latexextra bundle; English-only manuscripts can drop the CJK entries.

Usage

YouRA can be run in two modes. The benchmark runs in this repository use the unattended Python launcher for reproducibility. Human users can also open Claude Code in YOURA/ and use slash commands such as /phase0-brainstorm, /phase1-research, /phase2a-dialogue, and /hypothesis-loop to enter a specific phase directly and proceed interactively with the AI. See YOURA/README.md for the detailed YouRA workflow layout and phase-command summary.

Quick Start: Full Unattended Pipeline

Run the pipeline from the YOURA/ subdirectory. If you are at the repository root, enter it first. Include --enable-refine so Phase 6.5.1 is followed by the final refinement/PDF pass.

cd YOURA

# Text input
python .claude/hooks/run_total_youra.py "Weak supervision for image classification" \
    --research-folder docs/youra_research/20260304_scsl \
    --enable-refine

# Existing research idea file with a fixed output folder
python .claude/hooks/run_total_youra.py docs/research_idea.md \
    --research-folder docs/youra_research/20260304_scsl \
    --enable-refine

Interactive Start for Very Short Research Topics

The commands above run the end-to-end pipeline automatically. For human-guided use, avoid starting the full pipeline from an overly terse research topic such as weak supervision or LLM alignment. A short topic can make Phase 0 generate a too-simple idea and continue before you have shaped the research direction.

Instead, start Claude Code from YOURA/ and run the Phase 0 slash command first:

/phase0-brainstorm

Use Phase 0 to expand the topic into a concrete research question and check the generated docs/youra_research/<run>/00_brainstorm_session.md. After reviewing or editing that Phase 0 output, continue the automated pipeline from Phase 1:

python .claude/hooks/run_total_youra.py dummy \
    --resume-from phase1 \
    --research-folder docs/youra_research/<run> \
    --enable-refine

Claude Code also exposes phase-by-phase slash commands if you want to inspect or steer each phase interactively. See YOURA/README.md for the command summary.

The full pipeline runs:

Phase 0  Problem Scoping / Brainstorm
Phase 1  Literature Grounding
Phase 2A Hypothesis Dialogue
Phase 2B Verification Planning
Phase 2C Experiment Design
Phase 3  Implementation Planning
Phase 4  Coding & Validation
Phase 4.5 Hypothesis Synthesis
Phase 6  Manuscript Drafting
Phase 6.5 Adversarial Fact Review
Phase 6.5.1 Overleaf LaTeX + PDF Generation
Refine   Manuscript refinement

Runtime Configuration

Default run behavior can be edited in YOURA/.claude/hooks/auto_responder_config.yaml.

Setting	YAML Path	Meaning
Claude Code model	claude_model	Selects the Claude CLI execution model used by run_phase*.py, for example claude-sonnet-4-6, claude-opus-4-6, or claude-haiku-4-5-20251001. Empty/null uses the Claude CLI default.
Claude effort	claude_effort	Controls Claude extended thinking level: low, medium, high, or max.
Claude thinking switch	claude_thinking	false sets MAX_THINKING_TOKENS=0 for every Claude CLI subprocess, disabling extended thinking regardless of claude_effort; true (or absent) lets claude_effort / the CLI default apply. The shipped config uses claude_model: claude-sonnet-4-6, claude_effort: low, claude_thinking: false.
Reflection budget	pipeline_reflection.max_reflections	Default maximum number of reflection reroutes after MUST_WORK failures. -1 means unlimited; 0 disables reflection. CLI --max-reflections can be used per run.
Auto-responder model	openrouter.model	Model used by the OpenRouter-based auto-responder/controller when OpenRouter analysis is enabled.

Resume From The Middle

Use --resume-from when intermediate outputs already exist. Resume points other than phase0 require --research-folder. Keep --enable-refine on resumed commands so the final refinement pass runs after Phase 6.5.1.

# Resume from any supported phase by changing the --resume-from value
python .claude/hooks/run_total_youra.py dummy \
    --resume-from hypothesis-loop \
    --research-folder docs/youra_research/20260304_scsl \
    --enable-refine

Supported --resume-from values:

Value	Meaning
phase0, phase1, phase2a, phase2b	Restart the early pipeline at the selected phase.
hypothesis-loop	Skip early phases and resume Phase 2C -> 3 -> 4.
phase45, phase6, phase65, phase651	Resume the post-experiment manuscript pipeline.
refine	Run only the refinement pass on an existing Phase 6.5.1 output; requires --enable-refine.

phase5 is also accepted, but Phase 5 (baseline comparison) is not shipped: run_phase5.py does not exist and --enable-phase5 aborts, so without that flag --resume-from phase5 behaves the same as phase6.

Running the no-VSA Ablation Variant (YOURA_no_VSA/)

YOURA_no_VSA/ is a self-contained copy of the agent used for the persistence-versus-context ablation (the *_no_VSA lanes in results/README.md). It shares the full pipeline, launchers, and phase commands with YOURA/; the differences are:

• A guard_state_files.py PreToolUse hook blocks the task model from reading or writing the durable VSA files (verification_state.yaml and its checkpoints).
• ablation_state_manager.py supplies the equivalent state as prompt-visible context and maintains harness-only shadow copies under docs/youra_research/<run>/.ablation_shadow/, preserving identical stage transitions.
• ablation_audit.py verifies after each run that the task model made no reads or writes to the durable VSA files.

Setup mirrors YOURA/ — run the hook installer from inside YOURA_no_VSA/:

cd YOURA_no_VSA
python install_hooks.py --install-deps

Always pass --state-mode shadow. The flag defaults to normal (the environment variable YOURA_STATE_MODE=shadow changes that default); a run launched in normal mode behaves like the full system and cannot be used as an ablation arm.

cd YOURA_no_VSA
python .claude/hooks/run_total_youra.py tasks_youra/iclr2025_scsl.md \
    --research-folder docs/youra_research/20260304_scsl \
    --enable-refine \
    --state-mode shadow

All other launcher options (--resume-from, --max-reflections, the runtime configuration in .claude/hooks/auto_responder_config.yaml) work exactly as described above for YOURA/; keep --state-mode shadow on resumed commands too. The artifacts generated by this variant for the paper are bundled under results/generations/youra/<backbone>_no_VSA/.

Running the Controller-off Ablation Variant (YOURA_no_IC/)

YOURA_no_IC/ is a self-contained copy of the agent used for the independent-controller ablation (the *_no_IC lanes in results/README.md). The independent GPT-5.2 controller route is disabled: controller-owned lifecycle and recovery decisions, plus debate and review moderation, run from predefined static prompts. Concretely, phase_auto_responder.py gains a responder_mode: "fixed" path (enabled in .claude/hooks/auto_responder_config.yaml) that replaces the GPT-5.2 stop-hook decision with the deterministic phase_output_verifier.py checks plus a fixed resume prompt, and run_phase2a.py runs the hypothesis debate as a self-play loop inside the execution session instead of through the external orchestrate_exchange.py moderator, so debate itself is conducted by the execution model. The persistent VSA, MCP tools, Reflective Memory, and stage graph are unchanged.

Setup and launch are identical to YOURA/ — no extra flag is needed; the substitution lives in the hook code and config. The artifacts generated by this variant are bundled under results/generations/youra/<backbone>_no_IC/.

Running the Combined no-VSA + Controller-off Variant (YOURA_no_VSA_no_IC/)

YOURA_no_VSA_no_IC/ is a self-contained copy of the agent that applies both ablations above at once (the sonnet45_no_VSA_no_IC, sonnet46_no_VSA_no_IC, and opus45_no_VSA_no_IC lanes in results/README.md):

• The no-VSA hooks from YOURA_no_VSA/ (guard_state_files.py, ablation_state_manager.py, ablation_audit.py) block the task model from the durable VSA files and supply the equivalent state as prompt-visible context.
• The controller-off substitution from YOURA_no_IC/ (phase_auto_responder.py with responder_mode: "fixed" in .claude/hooks/auto_responder_config.yaml, and the self-play debate in run_phase2a.py) replaces the GPT-5.2 controller route with predefined static prompts; the external LLM is additionally hard-disabled in the phase configs.

Setup mirrors YOURA_no_VSA/, and --state-mode shadow is required for the same reason as there:

cd YOURA_no_VSA_no_IC
python install_hooks.py --install-deps
python .claude/hooks/run_total_youra.py tasks_youra/iclr2025_scsl.md \
    --research-folder docs/youra_research/20260304_scsl \
    --enable-refine \
    --state-mode shadow

The artifacts generated by this variant are bundled under results/generations/youra/<backbone>_no_VSA_no_IC/.

These four agent folders (YOURA/, YOURA_no_VSA/, YOURA_no_IC/, YOURA_no_VSA_no_IC/) are the only ones bundled. The *_no_mcp, *_no_reflection, and the three- and four-way combined lanes under results/ ship their generated artifacts and scores only; see results/README.md for their definitions.

MCP Tool Stack

Claude Code is the execution host and hook surface. Serena and Archon are the main recommended MCP-backed memory/tool layers; the remaining MCP services are optional and can be found in the Smithery server directory. The per-phase server requirements are listed in YOURA/bmad-custom-src/custom/modules/youra-research/mcp-phase-config.yaml; they are enforced by the workflow prompts, not by the Python launchers, so a missing server degrades tool grounding rather than aborting a run.

Tool	Status	Usage in YouRA
Serena	Strongly recommended (declared required by the phase workflows)	Code-aware symbol navigation/editing and Reflective Memory failure-context recording.
Archon	Recommended (declared required by the phase workflows)	Sequential Memory, RAG-backed knowledge base, implementation-pattern storage, and task lifecycle management.
Clear Thought, Exa, Semantic Scholar	Optional	Available through the Smithery server directory if you want structured reasoning, web evidence search, or literature search integrations.

Modified MLR-Bench Evaluation Utilities

This repository includes a modified mlrbench-youra package under src/mlrbench. It is used for evaluation, not for launching the YouRA research loop. The package provides unified judge wrappers for YouRA, AI Scientist V2, and MLR-Agent outputs; install it once from the repository root with pip install -e ..

→ See src/README.md for setup, the single-task run_eval runner, supported systems, output locations, and the experiments/-folder builder for new YouRA runs.

The package also carries the idea/proposal-stage pipeline behind the paper's stagewise comparison (Tables 13-14): src/mlrbench/evals/youra/generate_*.py regenerate idea.md/proposal.md/related_work.md from the source documents bundled under results/generations/idea_and_proposal/, and the review_idea.py/review_proposal.py scripts in evals/{youra,mlragent}/ re-run the stage judge reviews over those artifacts. → See results/README.md for the exact order, --verify modes, and judge notes.

Experiment Reproduction

All table- and figure-regeneration scripts, the pairwise LLM-as-judge harness, the data-provenance (Real/Synthetic/Fabricated) diagnostics, the ablation-study score statistics, and the human validation of hallucination flags live under analysis/. Install the mlrbench-youra package first (pip install -e .). The data-provenance diagnostic can be re-run over the bundled results/generations/ main lanes (three systems x three backbones x ten tasks) with a single command, python analysis/Data_type_fabrication_analysis/run_generations_data_type.py --engine both, which drives both the Claude (claude-opus-4-6) and Codex (gpt-5.4) analyzers.

→ See analysis/README.md for the per-folder guide and copy-paste commands that reproduce each paper artifact from the bundled evaluation outputs.

Project Structure

Your-Research-Agent/
+-- README.md                         # Repository-level overview (this file)
+-- overview.png                      # Main lifecycle overview figure
+-- pyproject.toml                    # Root evaluation package metadata
+-- YOURA/                            # The YouRA agent itself (see YOURA/README.md)
|   +-- install_hooks.py              # Claude Code hook installer
|   +-- setup_tex.py                  # TinyTeX + LaTeX package installer (PDF generation)
|   +-- requirements.txt              # Hook runtime dependencies
|   +-- tasks_youra/                  # Research task prompts
|   +-- .claude/
|   |   +-- commands/                 # Stage-specific Claude commands
|   |   +-- hooks/                    # Pipeline launchers, responder, verifier
|   |   +-- agents/                   # Phase-specialized agents
|   |   +-- prompts/                  # Phase responder prompts
|   |   +-- skills/                   # Tool/reasoning skill wrappers
|   +-- bmad-custom-src/              # YouRA BMAD module sources
+-- YOURA_no_VSA/                     # no-VSA ablation variant of the agent (see "Running the no-VSA Ablation Variant")
+-- YOURA_no_IC/                      # controller-off ablation variant of the agent (see "Running the Controller-off Ablation Variant")
+-- YOURA_no_VSA_no_IC/               # combined no-VSA + controller-off variant (see "Running the Combined no-VSA + Controller-off Variant")
+-- analysis/                         # Reproduction scripts (see analysis/README.md)
+-- results/                          # Evaluation outputs + generated artifacts (see results/README.md)
+-- src/
    +-- mlrbench/                     # Evaluation utilities (see src/README.md)

Each top-level folder has its own README with the full details:

Folder	README	Contents
YOURA/	YOURA/README.md	YouRA workflow layout, phase commands, and hook runtime.
analysis/	analysis/README.md	Scripts that reproduce every table and figure in the paper, incl. the ablation-study score statistics.
results/	results/README.md	Layout of the bundled evaluation scores (incl. youra_ablation_study) and generated research artifacts (incl. the ablation control lanes).
src/	src/README.md	The mlrbench-youra evaluation package and its CLI runners.

Results Format

Primary generated research artifacts are saved under YOURA/docs/youra_research/ and mirrored in benchmark-generation results under results/generations/youra/<lane>/<task>/docs/youra_research/. The final refined manuscript of each run is paper/refinement/06_paper_refinement.md.

→ See results/README.md for the full per-run layout, the evaluation-score folder structure, the ablation lanes, and the release anonymization notes.

Safety and Reliability Mechanisms

Mechanism	Purpose
Phase output verification	Each launcher checks required files and structured fields before advancing.
Auto-responder stop control	The hook router either approves stop, emits a resume prompt, or escalates MUST_STOP.
VSA gates	MUST_WORK and SHOULD_WORK gates determine whether failures trigger repair, redesign, reset, or limitation recording.
Independent fact review	Later manuscript stages check generated claims against VSA, stage artifacts, results, logs, and code.

Citation

@inproceedings{woo2026youra,
  title={YouRA: A Persistent-State Architecture for Evidence-Traceable Autonomous Research Agents},
  author={Woo, Yoonkyu and Lee, Woojin and Huang, Jin-xia},
  booktitle={Proceedings of AACL-IJCNLP 2026},
  year={2026},
  eprint={2610.01097},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2610.01097},
  note={Yoonkyu Woo and Woojin Lee contributed equally}
}

License

This repository is released under the PolyForm Noncommercial License 1.0.0. MLR-Bench tasks, reference papers, datasets, and starter code remain under their original upstream licenses; generated artifacts (papers, code, and logs under results/) are released under the same terms as this repository and inherit applicable upstream obligations.