Learning a Pessimistic Reward Model in RLHF

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Xu, Yinglun, Kang, Hangoo, Suresh, Tarun, Wan, Yuxuan, Singh, Gagandeep
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913859997007872
author Xu, Yinglun
Kang, Hangoo
Suresh, Tarun
Wan, Yuxuan
Singh, Gagandeep
author_facet Xu, Yinglun
Kang, Hangoo
Suresh, Tarun
Wan, Yuxuan
Singh, Gagandeep
contents This work proposes `PET', a novel pessimistic reward fine-tuning method, to learn a pessimistic reward model robust against reward hacking in offline reinforcement learning from human feedback (RLHF). Traditional reward modeling techniques in RLHF train an imperfect reward model, on which a KL regularization plays a pivotal role in mitigating reward hacking when optimizing a policy. Such an intuition-based method still suffers from reward hacking, and the policies with large KL divergence from the dataset distribution are excluded during learning. In contrast, we show that when optimizing a policy on a pessimistic reward model fine-tuned through PET, reward hacking can be prevented without relying on any regularization. We test our methods on the standard TL;DR summarization dataset. We find that one can learn a high-quality policy on our pessimistic reward without using any regularization. Such a policy has a high KL divergence from the dataset distribution while having high performance in practice. In summary, our work shows the feasibility of learning a pessimistic reward model against reward hacking. The agent can greedily search for the policy with a high pessimistic reward without suffering from reward hacking.
format Preprint
id arxiv_https___arxiv_org_abs_2505_20556
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Learning a Pessimistic Reward Model in RLHF
Xu, Yinglun
Kang, Hangoo
Suresh, Tarun
Wan, Yuxuan
Singh, Gagandeep
Machine Learning
This work proposes `PET', a novel pessimistic reward fine-tuning method, to learn a pessimistic reward model robust against reward hacking in offline reinforcement learning from human feedback (RLHF). Traditional reward modeling techniques in RLHF train an imperfect reward model, on which a KL regularization plays a pivotal role in mitigating reward hacking when optimizing a policy. Such an intuition-based method still suffers from reward hacking, and the policies with large KL divergence from the dataset distribution are excluded during learning. In contrast, we show that when optimizing a policy on a pessimistic reward model fine-tuned through PET, reward hacking can be prevented without relying on any regularization. We test our methods on the standard TL;DR summarization dataset. We find that one can learn a high-quality policy on our pessimistic reward without using any regularization. Such a policy has a high KL divergence from the dataset distribution while having high performance in practice. In summary, our work shows the feasibility of learning a pessimistic reward model against reward hacking. The agent can greedily search for the policy with a high pessimistic reward without suffering from reward hacking.
title Learning a Pessimistic Reward Model in RLHF
topic Machine Learning
url https://arxiv.org/abs/2505.20556