PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Jian, Ai, Ruan, Jingqing, Ma, Xing, Zhang, Xiaoyun, Li, Dailin, Zhang, Weipeng, Zeng, Ke, Cai, Xunliang
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913044092682240
author Jian, Ai
Ruan, Jingqing
Ma, Xing
Zhang, Xiaoyun
Li, Dailin
Zhang, Weipeng
Zeng, Ke
Cai, Xunliang
author_facet Jian, Ai
Ruan, Jingqing
Ma, Xing
Zhang, Xiaoyun
Li, Dailin
Zhang, Weipeng
Zeng, Ke
Cai, Xunliang
contents Reward models (RMs) are central to reinforcement learning from human feedback (RLHF), providing the critical supervision signals that align large language models (LLMs) with human preferences. Generative reward models (GRMs) provide greater interpretability than traditional scalar RMs, but they come with a critical trade-off: pairwise methods are hindered by a training-inference mismatch, while pointwise methods require expensive absolute annotations. To bridge this gap, we propose the Preference-aware Task-adaptive Reward Model (PaTaRM). Unlike prior approaches, PaTaRM enables robust pointwise training using readily available pairwise data via a novel Preference-Aware Reward (PAR) mechanism, eliminating the need for explicit rating labels. Furthermore, it incorporates a Task-Adaptive Rubric system that dynamically generates instance-specific criteria for precise evaluation. Extensive experiments demonstrate that PATRM achieves a 8.7% average improvement on RewardBench and RMBench across Qwen3-8B/14B models. Crucially, it boosts downstream RLHF performance by an average relative improvement of 13.6% across IFEval and InFoBench, validating its effectiveness for policy alignment. Our code is available at https://github.com/JaneEyre0530/PaTaRM.
format Preprint
id arxiv_https___arxiv_org_abs_2510_24235
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling
Jian, Ai
Ruan, Jingqing
Ma, Xing
Zhang, Xiaoyun
Li, Dailin
Zhang, Weipeng
Zeng, Ke
Cai, Xunliang
Machine Learning
Artificial Intelligence
Reward models (RMs) are central to reinforcement learning from human feedback (RLHF), providing the critical supervision signals that align large language models (LLMs) with human preferences. Generative reward models (GRMs) provide greater interpretability than traditional scalar RMs, but they come with a critical trade-off: pairwise methods are hindered by a training-inference mismatch, while pointwise methods require expensive absolute annotations. To bridge this gap, we propose the Preference-aware Task-adaptive Reward Model (PaTaRM). Unlike prior approaches, PaTaRM enables robust pointwise training using readily available pairwise data via a novel Preference-Aware Reward (PAR) mechanism, eliminating the need for explicit rating labels. Furthermore, it incorporates a Task-Adaptive Rubric system that dynamically generates instance-specific criteria for precise evaluation. Extensive experiments demonstrate that PATRM achieves a 8.7% average improvement on RewardBench and RMBench across Qwen3-8B/14B models. Crucially, it boosts downstream RLHF performance by an average relative improvement of 13.6% across IFEval and InFoBench, validating its effectiveness for policy alignment. Our code is available at https://github.com/JaneEyre0530/PaTaRM.
title PaTaRM: Bridging Pairwise and Pointwise Signals via Preference-Aware Task-Adaptive Reward Modeling
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2510.24235