Statistical Early Stopping for Reasoning Models

While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes overthink, generating unnecessary reasoning steps, particularly under uncertainty, given ill-posed or ambiguous queries. We introduce statistically principled early stopping methods that monitor uncertainty signals during generation to mitigate this issue. Our first approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. Our second approach is parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. We conduct empirical evaluations on reasoning tasks across several domains and models. Our results indicate that uncertainty-aware early stopping can improve both efficiency and reliability in LLM reasoning. The performance varies across domains, and we observe especially significant gains for math reasoning.

Language models sometimes overthink. When faced with ill-posed or ambiguous questions, instead of asking for clarification, they may generate unnecessarily long reasoning text. This wastes computational resources and can even lead to incorrectly guessed answers. Our goal is to help models stop reasoning when they are very uncertain. To achieve this, we identify uncertainty keywords in the model’s reasoning text, such as phrases like “cannot determine,” “perhaps intended,” or “missing information.” We then develop statistical methods that monitor the frequency of these uncertainty signals and decide when the model should stop reasoning. Our approach can potentially reduce computational cost. Rather than wasting resources on unnecessary reasoning, the model can stop early and, when appropriate, ask for clarification instead of making unreliable guesses.