TrajSelector : Harnessing Latent Representations for Efficient and Effective Best-of- N N in Large Reasoning Model

Bin Yu, Xinming Wang, Shijie Lian, Haotian Li, Changti Wu, Ruina Hu, Bailing Wang, Yuliang Wei, Kai Chen · 2025

Read TrajSelector : Harnessing Latent Representations for Efficient and Effective Best-of- N N in Large Reasoning Model by Bin Yu, Xinming Wang, Shijie Lian, Haotian Li, Changti Wu, Ruina Hu, Bailing Wang, Yuliang Wei, Kai Chen (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-2510-16449.txt.

Introduction

Bin Yu leads eight co-authors in a 2025 study on picking the best answer from a reasoning model. One way to boost scores is to make many tries and pick the best. That pick is often slow and costly. Their tool, TrajSelector, scores each try step by step. It reads hidden states inside the model that made the tries.

A small checker, with well under a billion weights, gives the step scores. Training needs no costly step-by-step human labels at all. In tests with 32 tries, it beat plain vote counts by a clear margin. It also beat older score models while costing less to run. Cheap picks make test-time scaling far more useful.

Catalog record

CREATORBin Yu, Xinming Wang, Shijie Lian, Haotian Li, Changti Wu, Ruina Hu, Bailing Wang, Yuliang Wei, Kai Chen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE37 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-16449

Browse the full library · About the Tower of Babel