MetaRM: Shifted Distributions Alignment via Meta-Learning

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Dou, Shihan, Liu, Yan, Zhou, Enyu, Li, Tianlong, Jia, Haoxiang, Xiong, Limao, Zhao, Xin, Ye, Junjie, Zheng, Rui, Gui, Tao, Zhang, Qi, Huang, Xuanjing
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866909186314469376
author Dou, Shihan
Liu, Yan
Zhou, Enyu
Li, Tianlong
Jia, Haoxiang
Xiong, Limao
Zhao, Xin
Ye, Junjie
Zheng, Rui
Gui, Tao
Zhang, Qi
Huang, Xuanjing
author_facet Dou, Shihan
Liu, Yan
Zhou, Enyu
Li, Tianlong
Jia, Haoxiang
Xiong, Limao
Zhao, Xin
Ye, Junjie
Zheng, Rui
Gui, Tao
Zhang, Qi
Huang, Xuanjing
contents The success of Reinforcement Learning from Human Feedback (RLHF) in language model alignment is critically dependent on the capability of the reward model (RM). However, as the training process progresses, the output distribution of the policy model shifts, leading to the RM's reduced ability to distinguish between responses. This issue is further compounded when the RM, trained on a specific data distribution, struggles to generalize to examples outside of that distribution. These two issues can be united as a challenge posed by the shifted distribution of the environment. To surmount this challenge, we introduce MetaRM, a method leveraging meta-learning to align the RM with the shifted environment distribution. MetaRM is designed to train the RM by minimizing data loss, particularly for data that can improve the differentiation ability to examples of the shifted target distribution. Extensive experiments demonstrate that MetaRM significantly improves the RM's distinguishing ability in iterative RLHF optimization, and also provides the capacity to identify subtle differences in out-of-distribution samples.
format Preprint
id arxiv_https___arxiv_org_abs_2405_00438
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle MetaRM: Shifted Distributions Alignment via Meta-Learning
Dou, Shihan
Liu, Yan
Zhou, Enyu
Li, Tianlong
Jia, Haoxiang
Xiong, Limao
Zhao, Xin
Ye, Junjie
Zheng, Rui
Gui, Tao
Zhang, Qi
Huang, Xuanjing
Machine Learning
Computation and Language
The success of Reinforcement Learning from Human Feedback (RLHF) in language model alignment is critically dependent on the capability of the reward model (RM). However, as the training process progresses, the output distribution of the policy model shifts, leading to the RM's reduced ability to distinguish between responses. This issue is further compounded when the RM, trained on a specific data distribution, struggles to generalize to examples outside of that distribution. These two issues can be united as a challenge posed by the shifted distribution of the environment. To surmount this challenge, we introduce MetaRM, a method leveraging meta-learning to align the RM with the shifted environment distribution. MetaRM is designed to train the RM by minimizing data loss, particularly for data that can improve the differentiation ability to examples of the shifted target distribution. Extensive experiments demonstrate that MetaRM significantly improves the RM's distinguishing ability in iterative RLHF optimization, and also provides the capacity to identify subtle differences in out-of-distribution samples.
title MetaRM: Shifted Distributions Alignment via Meta-Learning
topic Machine Learning
Computation and Language
url https://arxiv.org/abs/2405.00438