Saved in:
Bibliographic Details
Main Authors: Feng, Shengyu, He, Yun, Ma, Shuang, Li, Beibin, Xiong, Yuanhao, Li, Songlin, Mandyam, Karishma, Katz-Samuels, Julian, Bi, Shengjie, Yu, Licheng, Zhang, Hejia, Sankararaman, Karthik Abinav, Fang, Han, Yang, Yiming, Faruqui, Manaal
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2510.15242
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912957050388480
author Feng, Shengyu
He, Yun
Ma, Shuang
Li, Beibin
Xiong, Yuanhao
Li, Songlin
Mandyam, Karishma
Katz-Samuels, Julian
Bi, Shengjie
Yu, Licheng
Zhang, Hejia
Sankararaman, Karthik Abinav
Fang, Han
Yang, Yiming
Faruqui, Manaal
author_facet Feng, Shengyu
He, Yun
Ma, Shuang
Li, Beibin
Xiong, Yuanhao
Li, Songlin
Mandyam, Karishma
Katz-Samuels, Julian
Bi, Shengjie
Yu, Licheng
Zhang, Hejia
Sankararaman, Karthik Abinav
Fang, Han
Yang, Yiming
Faruqui, Manaal
contents Reinforcement learning (RL) has recently proven effective at scaling chain-of-thought (CoT) reasoning in large language models for tasks with verifiable answers. However, extending RL-based thought training to more general non-verifiable tasks-where supervision is provided only through pairwise human preferences-remains challenging. Existing approaches typically apply RL objectives designed for verifiable rewards to preference-based settings in a heuristic manner. In this work, we show that introducing CoT reasoning into preference modeling fundamentally changes the structure of the Bradley-Terry (BT) likelihood, as the reasoning process must be treated as a latent variable. This results in a preference likelihood expressed as a ratio of expectations over stochastic generation trajectories, which cannot be optimized using Jensen-style bounds or standard RL objectives. To address this challenge, we derive a consistent Monte Carlo estimator for the gradient of the resulting likelihood, leading to Bradley-Terry Policy Optimization (BTPO). Empirically, BTPO enables stable and effective training of generative preference models with CoT reasoning, consistently outperforming prior heuristic approaches across multiple benchmarks and model scales.
format Preprint
id arxiv_https___arxiv_org_abs_2510_15242
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Bradley-Terry Policy Optimization for Generative Preference Modeling
Feng, Shengyu
He, Yun
Ma, Shuang
Li, Beibin
Xiong, Yuanhao
Li, Songlin
Mandyam, Karishma
Katz-Samuels, Julian
Bi, Shengjie
Yu, Licheng
Zhang, Hejia
Sankararaman, Karthik Abinav
Fang, Han
Yang, Yiming
Faruqui, Manaal
Machine Learning
Reinforcement learning (RL) has recently proven effective at scaling chain-of-thought (CoT) reasoning in large language models for tasks with verifiable answers. However, extending RL-based thought training to more general non-verifiable tasks-where supervision is provided only through pairwise human preferences-remains challenging. Existing approaches typically apply RL objectives designed for verifiable rewards to preference-based settings in a heuristic manner. In this work, we show that introducing CoT reasoning into preference modeling fundamentally changes the structure of the Bradley-Terry (BT) likelihood, as the reasoning process must be treated as a latent variable. This results in a preference likelihood expressed as a ratio of expectations over stochastic generation trajectories, which cannot be optimized using Jensen-style bounds or standard RL objectives. To address this challenge, we derive a consistent Monte Carlo estimator for the gradient of the resulting likelihood, leading to Bradley-Terry Policy Optimization (BTPO). Empirically, BTPO enables stable and effective training of generative preference models with CoT reasoning, consistently outperforming prior heuristic approaches across multiple benchmarks and model scales.
title Bradley-Terry Policy Optimization for Generative Preference Modeling
topic Machine Learning
url https://arxiv.org/abs/2510.15242