Statistical Early Stopping for Reasoning Models
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866912906133635072 |
|---|---|
| author | Xie, Yangxinyu Wang, Tao Mallick, Soham Sun, Yan Noarov, Georgy Yu, Mengxin Mallick, Tanwi Su, Weijie J. Dobriban, Edgar |
| author_facet | Xie, Yangxinyu Wang, Tao Mallick, Soham Sun, Yan Noarov, Georgy Yu, Mengxin Mallick, Tanwi Su, Weijie J. Dobriban, Edgar |
| contents | 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 parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. Our second approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. 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, and we observe especially significant gains for math reasoning. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_13935 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Statistical Early Stopping for Reasoning Models Xie, Yangxinyu Wang, Tao Mallick, Soham Sun, Yan Noarov, Georgy Yu, Mengxin Mallick, Tanwi Su, Weijie J. Dobriban, Edgar Artificial Intelligence Machine Learning 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 parametric: it models inter-arrival times of uncertainty keywords as a renewal process and applies sequential testing for stopping. Our second approach is nonparametric and provides finite-sample guarantees on the probability of halting too early on well-posed queries. 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, and we observe especially significant gains for math reasoning. |
| title | Statistical Early Stopping for Reasoning Models |
| topic | Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2602.13935 |