Reproduce “Lomekwi: Resource-Bounded Tool Discovery in LLM Agents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Roshan Klein-Seetharaman, Daniel Wang, Andrew Xu
The authors measure tool discovery separately from task success in settings where tool use is optional. In Lomekwi, they distinguish curiosity (finding ingredients), recognition (choosing to build), and efficiency (succeeding after building), with grind accounting for wins without a tool. In Recognition MCP, Qwen3.5 2B to 27B oracle discovery falls from 0.29 to 0.02, while overall win rate rises from 0.46 to 0.88.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.