Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning logo

Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning

Free

Enhancing LLM reasoning via MCTS and iterative preference learning

FreeFree tier
Type
Open Source

About Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning

This paper introduces a method to enhance the reasoning capabilities of Large Language Models (LLMs) by combining Monte Carlo Tree Search (MCTS) with iterative preference learning, inspired by AlphaZero. The approach uses MCTS's look-ahead ability to collect step-level preference data, breaking down instance-level rewards into finer granularity. It combines outcome validation with stepwise self-evaluation to improve intermediate step consistency. The policy is updated via Direct Preference Optimization (DPO) using the generated step-level preference data, with a theoretical justification for using on-policy sampling. Extensive evaluations on arithmetic and commonsense reasoning tasks (GSM8K, MATH, ARC-C) show significant accuracy improvements over supervised fine-tuning baselines (e.g., +5.9% on GSM8K, +5.8% on MATH, +15.8% on ARC-C with Mistral-7B). The paper also analyzes the trade-off between training and inference compute. The code is publicly available.

Key Features

Monte Carlo Tree Search (MCTS) for step-level preference data collection
Breakdown of instance-level rewards into granular step-level signals
Combination of outcome validation and stepwise self-evaluation
Direct Preference Optimization (DPO) for policy updates
On-policy sampling for effective self-improvement
Theoretical analysis of training and inference compute trade-offs

Pros & Cons

Pros
  • Achieves significant accuracy improvements on multiple reasoning benchmarks
  • Uses fine-grained step-level preference signals for better learning
  • Combines outcome validation with self-evaluation for consistency
  • Open-source code available
Cons
  • May require substantial computational resources due to MCTS look-ahead
  • Effectiveness depends on careful tuning of MCTS and preference learning parameters

Best For

Arithmetic reasoning tasks (e.g., GSM8K, MATH)Commonsense reasoning tasks (e.g., ARC-C)General LLM reasoning enhancement

FAQ

What is the core idea of this method?
It uses Monte Carlo Tree Search to collect step-level preference data, then applies Direct Preference Optimization to iteratively improve the LLM policy, analogous to AlphaZero's approach.
What types of tasks were evaluated?
The method was evaluated on arithmetic reasoning (GSM8K, MATH) and commonsense reasoning (ARC-C) tasks.
Is the code available?
Yes, the code is publicly available at the URL provided in the paper.