Rethinking Reward Model Evaluation: Are We Barking up the Wrong Tree?

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Wen, Xueru, Lou, Jie, Lu, Yaojie, Lin, Hongyu, Yu, Xing, Lu, Xinyu, He, Ben, Han, Xianpei, Zhang, Debing, Sun, Le
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910826868244480
author Wen, Xueru
Lou, Jie
Lu, Yaojie
Lin, Hongyu
Yu, Xing
Lu, Xinyu
He, Ben
Han, Xianpei
Zhang, Debing
Sun, Le
author_facet Wen, Xueru
Lou, Jie
Lu, Yaojie
Lin, Hongyu
Yu, Xing
Lu, Xinyu
He, Ben
Han, Xianpei
Zhang, Debing
Sun, Le
contents Reward Models (RMs) are crucial for aligning language models with human preferences. Currently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data. Although this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored. In this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance. Our findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance. Moreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. Through the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization. This underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.
format Preprint
id arxiv_https___arxiv_org_abs_2410_05584
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Rethinking Reward Model Evaluation: Are We Barking up the Wrong Tree?
Wen, Xueru
Lou, Jie
Lu, Yaojie
Lin, Hongyu
Yu, Xing
Lu, Xinyu
He, Ben
Han, Xianpei
Zhang, Debing
Sun, Le
Machine Learning
Artificial Intelligence
Computation and Language
Reward Models (RMs) are crucial for aligning language models with human preferences. Currently, the evaluation of RMs depends on measuring accuracy against a validation set of manually annotated preference data. Although this method is straightforward and widely adopted, the relationship between RM accuracy and downstream policy performance remains under-explored. In this work, we conduct experiments in a synthetic setting to investigate how differences in RM measured by accuracy translate into gaps in optimized policy performance. Our findings reveal that while there is a weak positive correlation between accuracy and downstream performance, policies optimized towards RMs with similar accuracy can exhibit quite different performance. Moreover, we discover that the way of measuring accuracy significantly impacts its ability to predict the final policy performance. Through the lens of the Regressional Goodhart effect, we recognize that accuracy, when used for measuring RM quality, can fail to fully capture the potential RM overoptimization. This underscores the inadequacy of relying solely on accuracy to reflect their impact on policy optimization.
title Rethinking Reward Model Evaluation: Are We Barking up the Wrong Tree?
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2410.05584