Improving Value-based Process Verifier via Structural Prior Injection

Zetian sun, Dongfang Li, Baotian Hu, Jun Yu, Min Zhang · 2025

Read Improving Value-based Process Verifier via Structural Prior Injection by Zetian sun, Dongfang Li, Baotian Hu, Jun Yu, Min Zhang (2025) 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-2502-17498.txt.

Description

Zetian Sun and co-authors improve how process verifiers judge steps in model reasoning, in work from 2025. Verifiers often estimate the value of a step by sampling many endings, but small samples add noise and error.

The team add a structural prior by turning a single value into the mean of a fixed distribution. Sampling error then becomes a gap between two distributions, which training can shrink. Tests on best-of-many and beam search tasks show gains of about one to two points at little cost. The choice of prior mattered a lot, even when best answers matched.

Catalog record

CREATORZetian sun, Dongfang Li, Baotian Hu, Jun Yu, Min Zhang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE46 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-17498

Browse the full library · About the Tower of Babel