Believe Your Model: Distribution-Guided Confidence Calibration

Xizhong Yang, Haotian Zhang, Huiming Wang, Mofei Song, [6pt] Southeast University · 2026

Read Believe Your Model: Distribution-Guided Confidence Calibration by Xizhong Yang, Haotian Zhang, Huiming Wang, Mofei Song, [6pt] Southeast University (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-2603-03872.txt.

Description

Xizhong Yang and others propose DistriVoting, a method for choosing the best answer among several candidates generated by large reasoning models. Test-time scaling raises accuracy by sampling many responses, but picking the right one is the hard part. Internal confidence scores partly track correctness and correlate with accuracy in distribution, yet earlier methods never used that distributional information to guide selection.

DistriVoting first splits the mixed confidence distribution into positive and negative components with a Gaussian mixture model, then applies a rejection filter based on samples from each to reduce their overlap. A companion technique, SelfStepConf, uses step-level confidence to adjust the inference process itself and widen the separation. Across many models and five benchmarks, the method beats prior state-of-the-art approaches, showing that distributional priors make confidence-based voting markedly more reliable.

Catalog record

CREATORXizhong Yang, Haotian Zhang, Huiming Wang, Mofei Song, [6pt] Southeast University
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE136576
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.03872

Browse the full library · About the Tower of Babel