Saved in:
Bibliographic Details
Main Authors: Fang, Feiteng, Chen, Dingwei, Huang, Xiang, Lin, Ting-En, Wu, Yuchuan, Liu, Xiong, Ye, Xinge, Liu, Ziqiang, Zhang, Haonan, Zhu, Liang, Alinejad-Rokny, Hamid, Yang, Min, Li, Yongbin
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2505.23923
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912810050519040
author Fang, Feiteng
Chen, Dingwei
Huang, Xiang
Lin, Ting-En
Wu, Yuchuan
Liu, Xiong
Ye, Xinge
Liu, Ziqiang
Zhang, Haonan
Zhu, Liang
Alinejad-Rokny, Hamid
Yang, Min
Li, Yongbin
author_facet Fang, Feiteng
Chen, Dingwei
Huang, Xiang
Lin, Ting-En
Wu, Yuchuan
Liu, Xiong
Ye, Xinge
Liu, Ziqiang
Zhang, Haonan
Zhu, Liang
Alinejad-Rokny, Hamid
Yang, Min
Li, Yongbin
contents Currently, most reinforcement learning tasks focus on domains like mathematics and programming, where verification is relatively straightforward. However, in subjective tasks such as role-playing, alignment techniques struggle to make progress, primarily because subjective reward modeling using the Bradley-Terry model faces significant challenges when dealing with ambiguous preferences. To improve reward modeling in subjective tasks, this paper proposes AAM (\textbf{\underline{A}}ct-\textbf{\underline{A}}daptive \textbf{\underline{M}}argin), which enhances reward modeling by dynamically calibrating preference margins using the model's internal parameter knowledge. We design two versions of AAM that efficiently generate contextually-appropriate preference gaps without additional human annotation. This approach fundamentally improves how reward models handle subjective rewards by better integrating generative understanding with preference scoring. To validate AAM's effectiveness in subjective reward modeling, we conduct evaluations on RewardBench, JudgeBench, and challenging role-playing tasks. Results show that AAM significantly improves subjective reward modeling performance, enhancing Bradley-Terry reward models by 2.95\% in general tasks and 4.85\% in subjective role-playing tasks. Furthermore, reward models trained with AAM can help downstream alignment tasks achieve better results. Our test results show that applying rewards generated by AAM-Augmented RM to preference learning techniques (e.g., GRPO) achieves state-of-the-art results on CharacterEval and Charm. Code and dataset are available at https://github.com/calubkk/AAM.
format Preprint
id arxiv_https___arxiv_org_abs_2505_23923
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Act-Adaptive Margin: Dynamically Calibrating Reward Models for Subjective Ambiguity
Fang, Feiteng
Chen, Dingwei
Huang, Xiang
Lin, Ting-En
Wu, Yuchuan
Liu, Xiong
Ye, Xinge
Liu, Ziqiang
Zhang, Haonan
Zhu, Liang
Alinejad-Rokny, Hamid
Yang, Min
Li, Yongbin
Computation and Language
Artificial Intelligence
Currently, most reinforcement learning tasks focus on domains like mathematics and programming, where verification is relatively straightforward. However, in subjective tasks such as role-playing, alignment techniques struggle to make progress, primarily because subjective reward modeling using the Bradley-Terry model faces significant challenges when dealing with ambiguous preferences. To improve reward modeling in subjective tasks, this paper proposes AAM (\textbf{\underline{A}}ct-\textbf{\underline{A}}daptive \textbf{\underline{M}}argin), which enhances reward modeling by dynamically calibrating preference margins using the model's internal parameter knowledge. We design two versions of AAM that efficiently generate contextually-appropriate preference gaps without additional human annotation. This approach fundamentally improves how reward models handle subjective rewards by better integrating generative understanding with preference scoring. To validate AAM's effectiveness in subjective reward modeling, we conduct evaluations on RewardBench, JudgeBench, and challenging role-playing tasks. Results show that AAM significantly improves subjective reward modeling performance, enhancing Bradley-Terry reward models by 2.95\% in general tasks and 4.85\% in subjective role-playing tasks. Furthermore, reward models trained with AAM can help downstream alignment tasks achieve better results. Our test results show that applying rewards generated by AAM-Augmented RM to preference learning techniques (e.g., GRPO) achieves state-of-the-art results on CharacterEval and Charm. Code and dataset are available at https://github.com/calubkk/AAM.
title Act-Adaptive Margin: Dynamically Calibrating Reward Models for Subjective Ambiguity
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2505.23923