Reproduce “Q*: Improving Multi-step Reasoning for LLMs with Deliberative Planning” 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
Chaojie Wang, Yanchen Deng, Zhiyi Lyu, Liang Zeng, Jujie He, Shuicheng Yan, Bo An
Q* introduces a deliberative planning framework that enhances large language models' multi-step reasoning by integrating a learned Q-value model as a plug-and-play heuristic. This approach guides LLM generation through an A*-like search, significantly improving accuracy on complex math and code generation tasks without requiring fine-tuning of the base LLM.
Key insights
- Multi-step reasoning in LLMs can be effectively framed as a heuristic search problem, enabling deliberative planning beyond auto-regressive generation.
- A separate, lightweight Q-value model can serve as a general and effective "plug-and-play" heuristic to guide LLM decoding without modifying the base model.
- Combining process-based rewards (g-value) with learned future-oriented Q-value predictions (h-value) within an A*-like search significantly improves the selection of optimal intermediate reasoning steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.