LongRM: Revealing and Unlocking the Context Boundary of Reward Modeling

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Tang, Zecheng, Ji, Baibei, Qiu, Quantong, Wang, Haitian, Liang, Xiaobo, Li, Juntao, Zhang, Min
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866917057707114496
author Tang, Zecheng
Ji, Baibei
Qiu, Quantong
Wang, Haitian
Liang, Xiaobo
Li, Juntao
Zhang, Min
author_facet Tang, Zecheng
Ji, Baibei
Qiu, Quantong
Wang, Haitian
Liang, Xiaobo
Li, Juntao
Zhang, Min
contents Reward model (RM) plays a pivotal role in aligning large language model (LLM) with human preferences. As real-world applications increasingly involve long history trajectories, e.g., LLM agent, it becomes indispensable to evaluate whether a model's responses are not only high-quality but also grounded in and consistent with the provided context. Yet, current RMs remain confined to short-context settings and primarily focus on response-level attributes (e.g., safety or helpfulness), while largely neglecting the critical dimension of long context-response consistency. In this work, we introduce Long-RewardBench, a benchmark specifically designed for long-context RM evaluation, featuring both Pairwise Comparison and Best-of-N tasks. Our preliminary study reveals that even state-of-the-art generative RMs exhibit significant fragility in long-context scenarios, failing to maintain context-aware preference judgments. Motivated by the analysis of failure patterns observed in model outputs, we propose a general multi-stage training strategy that effectively scales arbitrary models into robust Long-context RMs (LongRMs). Experiments show that our approach not only substantially improves performance on long-context evaluation but also preserves strong short-context capability. Notably, our 8B LongRM outperforms much larger 70B-scale baselines and matches the performance of the proprietary Gemini 2.5 Pro model.
format Preprint
id arxiv_https___arxiv_org_abs_2510_06915
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle LongRM: Revealing and Unlocking the Context Boundary of Reward Modeling
Tang, Zecheng
Ji, Baibei
Qiu, Quantong
Wang, Haitian
Liang, Xiaobo
Li, Juntao
Zhang, Min
Computation and Language
Artificial Intelligence
Reward model (RM) plays a pivotal role in aligning large language model (LLM) with human preferences. As real-world applications increasingly involve long history trajectories, e.g., LLM agent, it becomes indispensable to evaluate whether a model's responses are not only high-quality but also grounded in and consistent with the provided context. Yet, current RMs remain confined to short-context settings and primarily focus on response-level attributes (e.g., safety or helpfulness), while largely neglecting the critical dimension of long context-response consistency. In this work, we introduce Long-RewardBench, a benchmark specifically designed for long-context RM evaluation, featuring both Pairwise Comparison and Best-of-N tasks. Our preliminary study reveals that even state-of-the-art generative RMs exhibit significant fragility in long-context scenarios, failing to maintain context-aware preference judgments. Motivated by the analysis of failure patterns observed in model outputs, we propose a general multi-stage training strategy that effectively scales arbitrary models into robust Long-context RMs (LongRMs). Experiments show that our approach not only substantially improves performance on long-context evaluation but also preserves strong short-context capability. Notably, our 8B LongRM outperforms much larger 70B-scale baselines and matches the performance of the proprietary Gemini 2.5 Pro model.
title LongRM: Revealing and Unlocking the Context Boundary of Reward Modeling
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2510.06915