Inference Time Optimization with Confidence Dynamics
Published in International Conference on Machine Learning (ICML 2026), 2026
Inference time optimization techniques, such as repeated sampling, have significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, the critical role of model uncertainty remains largely underexplored in these optimization strategies. In this paper, we investigate the dynamics of confidence along reasoning trajectories and for the first time reveal a surprising and unique pattern: correct answer traces tend to exhibit confidence improvement over time (positive confidence gain), while incorrect traces show attenuated or declining confidence as reasoning proceeds. Based on this observation, we propose Confidence Dynamic Gain (CDG) based voting, which incorporates how the confidence trajectory of the response evolves along the reasoning chain. Experiments across four open-source architectures (DeepSeek-R1, gpt-oss, Gemma-3, Qwen-QwQ) on the AIME24/25, HMMT25, and BRUMO25 benchmarks demonstrate that CDG yields a significant performance boost over baselines. These results demonstrate that our method provides a robust discriminative signal for improving answer selection in LLM reasoning. We also provide theoretical insights for this phenomenon. Code: https://github.com/Accenture/CDG.git
Recommended citation: @inproceedings{wang2026inference, title={Inference Time Optimization with Confidence Dynamics}, author={Wang, Yu and Liu, Minghao and Wang, Jiayun and Huang, Jinrui and Shah, Ankit and Wei, Wei}, booktitle={International Conference on Machine Learning (ICML)}, year={2026} }
Download Paper