Reproduce “Empirical-MCTS: Continuous Agent Evolution via Dual-Experience Monte Carlo Tree Search” 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
Hao Lu, Haoyuan Huang, Yulin Zhou, Chen Li, Ningxin Zhu
Developed by JianChengXingYun Technology Co., Ltd., Empirical-MCTS is a dual-experience framework that enhances Large Language Model reasoning by integrating Monte Carlo Tree Search with continuous, non-parametric learning. This approach achieved 73.3% on AIME25 and demonstrated improved cost-efficiency, allowing smaller LLMs to achieve competitive performance on complex reasoning tasks.
Key insights
- Combining dynamic, context-specific meta-prompt evolution with continuous, global experience accumulation enables LLMs to transition from stateless to stateful reasoning.
- Non-parametric updates to a global experience repository can serve as an effective policy prior, driving continuous self-improvement for LLM reasoning without requiring model retraining.
- Dynamic evaluation criteria and meta-prompt refinement, derived from pairwise comparisons of reasoning paths, are essential for leveraging accumulated knowledge effectively in complex tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.