West-of-N: Synthetic Preference Generation for Improved Reward Modeling

Alizée Pace, Jonathan Mallinson, Eric Malmi, Sebastian Krause, Aliaksei Severyn

Keywords: reinforcement learning from human feedback, reward modeling, preference data, self-training, large language models

TL;DR: West-of-N generates synthetic preference data for RLHF by self-training reward models. Performance gains are similar to collecting HF data.

Abstract: 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 improving 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.

Submission Number: 59