Evidence Information Gain for LLM Confidence Estimation: A Diffusion Decision Model-Inspired Approach

Estimating the confidence of Large Language Models (LLMs) remains a challenging problem, particularly when models generate complex reasoning processes. A prominent approach is to move beyond assigning a single confidence score to the entire reasoning chain and instead estimate trajectory-level confidence, capturing how confidence evolves throughout the reasoning process. In neuroscience, the Diffusion Decision Model (DDM) models decision-making as sequential evidence accumulation under noise, where a decision is reached when accumulated evidence crosses a decision boundary. Inspired by this perspective, we view the reasoning steps generated by an LLM as sequential evidence accumulated toward its final answer, and hypothesize that the contribution of reasoning steps to the final decision is informative for confidence estimation. To operationalize this idea, we first introduce a framework for quantifying the information gain of each evidence with respect to the final answer. We then derive novel features from the resulting evidence-accumulation patterns and use them to estimate the confidence of the model's final prediction. Across seven diverse benchmarks, our approach consistently outperforms existing confidence estimation methods, achieving meaningful improvements in AUROC while reducing ECE. Importantly, our method is applicable in both white-box and black-box settings, with the black-box variant achieving confidence estimation performance very close to that of the white-box setting.