Adaptive Uncertainty-Aware Tree Search for Robust Reasoning
Zeen Song, Zihao Ma, Wenwen Qiang, Changwen Zheng, Gang Hua · 2026
Read Adaptive Uncertainty-Aware Tree Search for Robust Reasoning by Zeen Song, Zihao Ma, Wenwen Qiang, Changwen Zheng, Gang Hua (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-06493.txt.
Description
Reasoning models that search at run time lean on reward models to score partial paths, but those scores turn unreliable off the training distribution. Zeen Song and 4 co-authors analyze that uncertainty problem and propose Uncertainty-Aware Tree Search. Standard search, they prove, piles up regret linearly when scores mislead, while uncertainty-aware search can keep regret sublinear.
Their method estimates scoring uncertainty with Monte Carlo dropout, which samples the reward model with random units dropped. A reinforcement learning controller then shifts compute toward paths where the score can be trusted and away from dubious ones. Experiments showed the approach reduced the harm of out-of-distribution scoring errors across reasoning tasks.
Catalog record
| CREATOR | Zeen Song, Zihao Ma, Wenwen Qiang, Changwen Zheng, Gang Hua |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 59 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06493