Reproduce “MCTS Is Counting the Wrong Thing: Learning Exploration Counts for LLM Reasoning” 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

Xuan Zhang, Zhijian Zhou, Zicheng Hu, Yan Feng, Hongyan Hao, Yuan Qi, Chao Qu

LPC-MCTS changes tree search for language-model reasoning by using learned pseudo-counts to estimate whether branches represent familiar strategies, rather than counting textual branches separately. A Coin-Flipping Network reads policy representations and adjusts exploration bonuses. With the same trained policy, LPC-MCTS inference raises mean@16 accuracy from 29.00 to 32.79 across five math benchmarks; this does not establish uniformly faster inference.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.