Reproduce “Action-Space Shaping for LLM Agents: Measuring and Mitigating Tool-Schema Bias” 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

Yinhong Liu, Zhili Tan, Zilin Wang, Zhijiang Guo

The authors test whether LLM agents perform consistently when tool schemas change while executable actions remain fixed. They map nine schema operators back to native actions and measure exact native-action match success. On 2,748 tasks, Qwen3-4B scores 77% with the native schema and 7% with schema discovery. Mixed-schema training raises discovery success to 75%, but evidence covers only tested variants.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.