Saved in:
Bibliographic Details
Main Authors: Pace, Alizée, Mallinson, Jonathan, Malmi, Eric, Krause, Sebastian, Severyn, Aliaksei
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2401.12086
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912084706459648
author Pace, Alizée
Mallinson, Jonathan
Malmi, Eric
Krause, Sebastian
Severyn, Aliaksei
author_facet Pace, Alizée
Mallinson, Jonathan
Malmi, Eric
Krause, Sebastian
Severyn, Aliaksei
contents The success of reinforcement learning from human feedback (RLHF) in language model alignment is strongly dependent on the quality of the underlying reward model. In this paper, we present a novel approach to improve reward model quality by generating synthetic preference data, thereby augmenting the training dataset with on-policy, high-quality preference pairs. Motivated by the promising results of Best-of-N sampling strategies in language model training, we extend their application to reward model training. This results in a self-training strategy to generate preference pairs by selecting the best and worst candidates in a pool of responses to a given query. Empirically, we find that this approach improves the performance of any reward model, with an effect comparable to the addition of a similar quantity of human preference data. This work opens up new avenues of research for improving RLHF for language model alignment, by offering synthetic preference generation as a solution to reward modeling challenges.
format Preprint
id arxiv_https___arxiv_org_abs_2401_12086
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle West-of-N: Synthetic Preferences for Self-Improving Reward Models
Pace, Alizée
Mallinson, Jonathan
Malmi, Eric
Krause, Sebastian
Severyn, Aliaksei
Computation and Language
Artificial Intelligence
Machine Learning
The success of reinforcement learning from human feedback (RLHF) in language model alignment is strongly dependent on the quality of the underlying reward model. In this paper, we present a novel approach to improve reward model quality by generating synthetic preference data, thereby augmenting the training dataset with on-policy, high-quality preference pairs. Motivated by the promising results of Best-of-N sampling strategies in language model training, we extend their application to reward model training. This results in a self-training strategy to generate preference pairs by selecting the best and worst candidates in a pool of responses to a given query. Empirically, we find that this approach improves the performance of any reward model, with an effect comparable to the addition of a similar quantity of human preference data. This work opens up new avenues of research for improving RLHF for language model alignment, by offering synthetic preference generation as a solution to reward modeling challenges.
title West-of-N: Synthetic Preferences for Self-Improving Reward Models
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2401.12086