Statistical Early Stopping for Reasoning Models

Yangxinyu Xie, Tao Wang, Soham Mallick, Yan Sun, Georgy Noarov, Mengxin Yu, Tanwi Mallick, Edgar Dobriban
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:139759-139785, 2026.

Abstract

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.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-xie26z, title = {Statistical Early Stopping for Reasoning Models}, author = {Xie, Yangxinyu and Wang, Tao and Mallick, Soham and Sun, Yan and Noarov, Georgy and Yu, Mengxin and Mallick, Tanwi and Dobriban, Edgar}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {139759--139785}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/xie26z/xie26z.pdf}, url = {https://proceedings.mlr.press/v306/xie26z.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T Statistical Early Stopping for Reasoning Models %A Yangxinyu Xie %A Tao Wang %A Soham Mallick %A Yan Sun %A Georgy Noarov %A Mengxin Yu %A Tanwi Mallick %A Edgar Dobriban %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-xie26z %I PMLR %P 139759--139785 %U https://proceedings.mlr.press/v306/xie26z.html %V 306 %X 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.
APA
Xie, Y., Wang, T., Mallick, S., Sun, Y., Noarov, G., Yu, M., Mallick, T. & Dobriban, E.. (2026). Statistical Early Stopping for Reasoning Models. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:139759-139785 Available from https://proceedings.mlr.press/v306/xie26z.html.

Related Material