Скачать или смотреть MITS: PMI-Guided Tree Search for LLM Reasoning

In this AI Research Roundup episode, Alex discusses the paper:
'MITS: Enhanced Tree Search Reasoning for LLMs via Pointwise Mutual Information'
This work tackles efficient, reliable multi-step reasoning in tree-search setups like Tree-of-Thought and MCTS by giving instant, quantitative scores to intermediate steps and exploring multiple paths without costly rollouts. MITS uses pointwise mutual information (PMI) to evaluate partial reasoning chains incrementally, normalizing by length to curb bias. It couples a generator and evaluator with adaptive sampling driven by token entropy, allocating more candidates to uncertain steps. Beam search prunes paths via cumulative PMI, and final answers are chosen with weighted voting across top chains.
Paper URL: https://arxiv.org/abs/2510.03632

#AI #MachineLearning #DeepLearning #LLM #Reasoning #TreeSearch #PMI #ChainOfThought