Value-Incentivized Preference Optimization: A Unified Approach to Online and Offline RLHF

TL;DR

This paper introduces VPO, a unified method for online and offline RLHF that incorporates uncertainty estimation into reward modeling, improving alignment of language models with human preferences.

Contribution

The paper proposes VPO, a novel approach that regularizes reward estimates with value functions, providing a unified, theoretically-grounded framework for RLHF in large language models.

Findings

VPO achieves competitive performance in text summarization and dialog tasks.

Theoretical guarantees match standard RL rates for both online and offline settings.

VPO simplifies the RLHF pipeline by integrating reward modeling and policy optimization.

Abstract

Reinforcement learning from human feedback (RLHF) has demonstrated great promise in aligning large language models (LLMs) with human preference. Depending on the availability of preference data, both online and offline RLHF are active areas of investigation. A key bottleneck is understanding how to incorporate uncertainty estimation in the reward function learned from the preference data for RLHF, regardless of how the preference data is collected. While the principles of optimism or pessimism under uncertainty are well-established in standard reinforcement learning (RL), a practically-implementable and theoretically-grounded form amenable to large language models is not yet available, as standard techniques for constructing confidence intervals become intractable under arbitrary policy parameterizations. In this paper, we introduce a unified approach to online and offline RLHF -- value-incentivized preference optimization (VPO) -- which regularizes the maximum-likelihood estimate of the reward function with the corresponding value function, modulated by a $\textit{sign}$ to indicate whether the optimism or pessimism is chosen. VPO also directly optimizes the policy with implicit reward modeling, and therefore shares a simpler RLHF pipeline similar to direct preference optimization. Theoretical guarantees of VPO are provided for both online and offline settings, matching the rates of their standard RL counterparts. Moreover, experiments on text summarization and dialog verify the practicality and effectiveness of VPO.