Reproduce “ChatCoT: Tool-Augmented Chain-of-Thought Reasoning on Chat-based Large Language Models” 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
Zhipeng Chen, Kun Zhou, Beichen Zhang, Zheng Gong, Wayne Xin Zhao, Ji-Rong Wen
ChatCoT unifies Chain-of-Thought reasoning with external tool use in large language models by framing their interaction as a multi-turn conversation. This framework achieved a 7.9% relative accuracy improvement on the MATH dataset and 59.2% accuracy on HotpotQA by enabling dynamic tool invocation within the reasoning process.
Key insights
- Modeling tool interaction as a natural multi-turn conversation unifies Chain-of-Thought reasoning and external tool usage, maintaining reasoning continuity.
- Leveraging the conversational capabilities of chat-based Large Language Models allows for dynamic, on-demand tool invocation without disrupting the reasoning flow.
- Pre-feeding the LLM with tool knowledge, retrieval-augmented task examples, and multi-turn reasoning formats is crucial for effective tool recognition and structured problem-solving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.