Mitigating Preference Hacking in Policy Optimization with Pessimism

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Gupta, Dhawal, Fisch, Adam, Dann, Christoph, Agarwal, Alekh
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910866793824256
author Gupta, Dhawal
Fisch, Adam
Dann, Christoph
Agarwal, Alekh
author_facet Gupta, Dhawal
Fisch, Adam
Dann, Christoph
Agarwal, Alekh
contents This work tackles the problem of overoptimization in reinforcement learning from human feedback (RLHF), a prevalent technique for aligning models with human preferences. RLHF relies on reward or preference models trained on \emph{fixed preference datasets}, and these models are unreliable when evaluated outside the support of this preference data, leading to the common reward or preference hacking phenomenon. We propose novel, pessimistic objectives for RLHF which are provably robust to overoptimization through the use of pessimism in the face of uncertainty, and design practical algorithms, P3O and PRPO, to optimize these objectives. Our approach is derived for the general preference optimization setting, but can be used with reward models as well. We evaluate P3O and PRPO on the tasks of fine-tuning language models for document summarization and creating helpful assistants, demonstrating remarkable resilience to overoptimization.
format Preprint
id arxiv_https___arxiv_org_abs_2503_06810
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Mitigating Preference Hacking in Policy Optimization with Pessimism
Gupta, Dhawal
Fisch, Adam
Dann, Christoph
Agarwal, Alekh
Machine Learning
Artificial Intelligence
This work tackles the problem of overoptimization in reinforcement learning from human feedback (RLHF), a prevalent technique for aligning models with human preferences. RLHF relies on reward or preference models trained on \emph{fixed preference datasets}, and these models are unreliable when evaluated outside the support of this preference data, leading to the common reward or preference hacking phenomenon. We propose novel, pessimistic objectives for RLHF which are provably robust to overoptimization through the use of pessimism in the face of uncertainty, and design practical algorithms, P3O and PRPO, to optimize these objectives. Our approach is derived for the general preference optimization setting, but can be used with reward models as well. We evaluate P3O and PRPO on the tasks of fine-tuning language models for document summarization and creating helpful assistants, demonstrating remarkable resilience to overoptimization.
title Mitigating Preference Hacking in Policy Optimization with Pessimism
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2503.06810