WeWorkBuddy

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

arXiv:2611.22301v1 Announce Type: new Abstract: Monte Carlo tree search (MCTS) has powered landmark successes in reinforcement learning, yet its gains in LLM reasoning can remain modest despite substantial search effort. We identify an overlooked bottleneck: what MCTS counts. Standard UCT assigns independent visit counts to textual branches, allowing different formulations of the same reasoning strategy to repeatedly receive exploration bonuses. Our theoretical analysis shows that this count fra

阅读原始来源
来源参考:MCTS Is Counting the Wrong Thing: Learning Exploration Counts for LLM Reasoning。暂未取得截图。
来源记录 · 预览待采集 · arxiv.org打开来源查看完整网页。

来源记录

项目分析

暂无五维分析,可先查看上方来源资料。