Lomekwi: Resource-Bounded Tool Discovery in LLM Agents

Code, environments, and run logs for the paper Lomekwi: Resource-Bounded Tool Discovery in LLM Agents (Klein-Seetharaman, Wang, Xu — Sea12 Technologies / Yale University).

Repository layout

• lomekwi.tex, lomekwi.bib — the paper.
• scripts/ — every runner, sweep, and plotting script that produces a number or figure in the paper, split by environment, plus a shared/ library (the per-episode relabeling/obfuscation protocol — three schemes, see Appendix D — and a multi-provider chat client) used by all of them. See scripts/README.md.
• runs/ — raw episode logs (episodes.jsonl) for the sweeps reported in the paper, one directory per sweep, grouped by environment (lomekwi/, deep_lomekwi/, oracle/).
• figures/ — the exact figure files cited in the paper.
• requirements.txt — Python dependencies.

Setup

pip install -r requirements.txt

Set whichever provider keys you need as environment variables (or in a .env file, loaded automatically): ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY (or GOOGLE_API_KEY); for local/open-weight models point VLLM_BASE_URL or OLLAMA_BASE_URL at a running server.

Reproducing results

All scripts run from the repo root with PYTHONPATH=., e.g.:

PYTHONPATH=. python -m scripts.lomekwi.run_grid_sweep
PYTHONPATH=. python -m scripts.lomekwi.plotters.plot_lomekwi_summary

See scripts/README.md for what each script reproduces. Sweeps write fresh logs under runs/; plotting scripts read those logs and write figures alongside them (or to figures/).

Citation

Klein-Seetharaman, R., Wang, D., and Xu, A.
Lomekwi: Resource-Bounded Tool Discovery in LLM Agents.