Revisiting Tree Search for LLMs: Gumbel and Sequential Halving for Budget-Scalable Reasoning

TL;DR

This paper introduces ReSCALE, a novel tree search method for LLM reasoning that maintains accuracy as search budgets increase, addressing prior scaling failures.

Contribution

ReSCALE replaces Dirichlet noise and PUCT with Gumbel sampling and Sequential Halving, improving reasoning accuracy without retraining models.

Findings

ReSCALE achieves 58.4% on GSM8K.

ReSCALE reaches 85.3% on Game24.

Sequential Halving is key to the performance gains.

Abstract

Neural tree search is a powerful decision-making algorithm widely used in complex domains such as game playing and model-based reinforcement learning. Recent work has applied AlphaZero-style tree search to enhance the reasoning capabilities of Large Language Models (LLMs) during inference, but we find that this approach suffers from a scaling failure: on GSM8K and Game24, accuracy drops as the search budget increases. In this paper, we present ReSCALE, an adaptation of Gumbel AlphaZero MCTS that replaces Dirichlet noise and PUCT selection with Gumbel sampling and Sequential Halving, restoring monotonic scaling without changes to the model or its training. ReSCALE reaches 58.4\% on GSM8K and 85.3\% on Game24 at budgets where the baseline degrades. Ablations confirm that Sequential Halving is the primary driver of the improvement.