Reproduce “Agents, Think Only When it Helps” 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
Min-Seon Kim, Seungdong Yoa, Sanghyu Yoon, Ye Seul Sim, Suhee Yoon, Dongmin Kim, Jung-Eun Kim
The paper examines when agent students need teacher reasoning traces alongside actions. It finds thought removal can preserve performance when actions expose relevant information, but can hurt when they do not. In MathQA, semantic action names raise Qwen2.5-3B-Instruct thought-drop accuracy from 10.7% to 16.7% and retention from 62.6% to 91.3%. These results support selective supervision on tested benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.