Adaptive self-consistency cuts reasoning samples in large language models
Adaptive Self-Consistency: From Black-Box Sampling to Distribution-Valued Feedback
Machine LearningArtificial Intelligence
Summary
When large language models give answers, they often try many possible reasoning paths and pick the most common final answer, but usually only one answer per path is used. The authors noticed that each answer actually comes with a hidden set of probabilities for all possible answers, not just one choice. They designed a way to use these probability details from each path to decide faster when enough evidence has been collected to trust the most likely answer. This method, called ASC-D, uses many fewer samples while still being confident in the chosen answer.
What this means in practice
- •For ai engineering teams: Reduce how many reasoning attempts large language models need to reach confident answers, saving computation time and cost.
- •For automated decision system developers: Improve reliability of systems that aggregate multiple model outputs by stopping early once consistent answers emerge with high confidence.
Authors
Jingkai Huang, Yunfan Zhang, Will Ma, Weihua Zhou, Zhengyuan Zhou
Abstract
Self-consistency samples many reasoning trajectories and aggregates their final answers, treating the LLM as a black box that returns one answer per trajectory. Yet the final answer of each trajectory is sampled from a softmax vector that is available from the model's log-probabilities. We refer to this as the grey-box setting in which each trajectory reveals this answer distribution rather than a single draw from it. We formulate efficient inference in this setting as sequential mode identification with distribution-valued observations: sample trajectories one at a time and stop as soon as the LLM's modal answer is identified at a prescribed confidence level. We characterize the asymptotic stopping rate of mode identification with distribution-valued observations exactly and show that it is never worse than the black-box rate. We then propose the ASC-D algorithm, a betting stopping rule that attains this asymptotic stopping rate. On MMLU-Redux, ASC-D uses $46.4$--$95.6\%$ fewer trajectories than answer-only adaptive self-consistency baselines and achieves the highest fixed-budget correct-certification rate across three open-source models.