Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Nguyen, Phuc Minh, Nguyen, Ngoc-Hieu, Nguyen, Duy H. M., Liu, Anji, Mai, An, Nguyen, Binh T., Sonntag, Daniel, Doan, Khoa D.
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2506.08681
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909645720780800
author Nguyen, Phuc Minh
Nguyen, Ngoc-Hieu
Nguyen, Duy H. M.
Liu, Anji
Mai, An
Nguyen, Binh T.
Sonntag, Daniel
Doan, Khoa D.
author_facet Nguyen, Phuc Minh
Nguyen, Ngoc-Hieu
Nguyen, Duy H. M.
Liu, Anji
Mai, An
Nguyen, Binh T.
Sonntag, Daniel
Doan, Khoa D.
contents Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO) have emerged as alternatives to the standard Reinforcement Learning from Human Feedback (RLHF) for aligning large language models (LLMs) with human values. However, these methods are more susceptible to over-optimization, in which the model drifts away from the reference policy, leading to degraded performance as training progresses. This paper proposes a novel importance-sampling approach to mitigate the over-optimization problem of offline DAAs. This approach, called (IS-DAAs), multiplies the DAA objective with an importance ratio that accounts for the reference policy distribution. IS-DAAs additionally avoid the high variance issue associated with importance sampling by clipping the importance ratio to a maximum value. Our extensive experiments demonstrate that IS-DAAs can effectively mitigate over-optimization, especially under low regularization strength, and achieve better performance than other methods designed to address this problem. Our implementations are provided publicly at this link.
format Preprint
id arxiv_https___arxiv_org_abs_2506_08681
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling
Nguyen, Phuc Minh
Nguyen, Ngoc-Hieu
Nguyen, Duy H. M.
Liu, Anji
Mai, An
Nguyen, Binh T.
Sonntag, Daniel
Doan, Khoa D.
Machine Learning
Direct Alignment Algorithms (DAAs) such as Direct Preference Optimization (DPO) have emerged as alternatives to the standard Reinforcement Learning from Human Feedback (RLHF) for aligning large language models (LLMs) with human values. However, these methods are more susceptible to over-optimization, in which the model drifts away from the reference policy, leading to degraded performance as training progresses. This paper proposes a novel importance-sampling approach to mitigate the over-optimization problem of offline DAAs. This approach, called (IS-DAAs), multiplies the DAA objective with an importance ratio that accounts for the reference policy distribution. IS-DAAs additionally avoid the high variance issue associated with importance sampling by clipping the importance ratio to a maximum value. Our extensive experiments demonstrate that IS-DAAs can effectively mitigate over-optimization, especially under low regularization strength, and achieve better performance than other methods designed to address this problem. Our implementations are provided publicly at this link.
title Mitigating Reward Over-optimization in Direct Alignment Algorithms with Importance Sampling
topic Machine Learning
url https://arxiv.org/abs/2506.08681