Reproduce “The Reasoning Trap: How Enhancing LLM Reasoning Amplifies Tool Hallucination” 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
Chenlong Yin, Zeyang Sha, Shiwen Cui, Changhua Meng, Zechao Li
The authors test whether reasoning-enabled agents abstain when a required tool is missing, using no-tool and irrelevant-distractor tasks. In a matched ablation on Qwen2.5-7B-Instruct trained on SynTool, direct tool-use reinforcement learning raises hallucination rates from 34.8% and 54.7% to 41.4% and 63.6%; requiring think-then-act raises them to 90.2% and 100.0%. The result is limited to tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.