Improving reasoning at inference time via uncertainty minimisation

TL;DR

This paper introduces a novel inference-time method for reasoning in large language models that maximizes internal self-certainty at each step, improving accuracy with fewer samples and better transferability across languages.

Contribution

The proposed approach frames reasoning as uncertainty minimization at the thought level, operating solely on internal model signals, and demonstrates superior performance over existing methods.

Findings

Outperforms greedy decoding and matches self-consistency with fewer samples

Transfers effectively across multiple languages and model sizes

Early reasoning decisions are predictive of final accuracy

Abstract

Large language models (LLMs) now exhibit strong multi-step reasoning abilities, but existing inference-time scaling methods remain computationally expensive, often relying on extensive sampling or external evaluators. We propose a principled strategy that frames reasoning as uncertainty minimisation and operates at the level of individual thoughts rather than tokens. Our method selects, at each reasoning step, the continuation that maximizes the model's self-certainty, a metric computed from its internal predictive distribution. This approach achieves significant improvement with a small number of samples, relies exclusively on model-internal signals, and applies to open-ended questions as opposed to methods like majority voting. Experiments on MATH500 and GSM8K across multiple model sizes demonstrate that thought-level self-certainty maximization consistently outperforms greedy decoding and matches or exceeds self-consistency under comparable token budgets. Cross-linguistic evaluations further indicate that the method transfers robustly beyond high-resource languages. Furthermore, analysis of self-certainty dynamics reveals that correct reasoning trajectories converge early to stable paths, suggesting that early decisions, likely associated with the planning of the reasoning process, are predictive of final accuracy. Building on this result, we show that self-certainty maximisation applied to the early steps can explain most of the performance gain and provide a simple yet efficient inference-time scaling method.