TGDPO: Harnessing Token-Level Reward Guidance for Enhancing Direct Preference Optimization

TL;DR

This paper introduces TGDPO, a method that leverages token-level reward guidance to improve Direct Preference Optimization, resulting in significant performance gains in aligning large language models with human preferences.

Contribution

The work decomposes sequence-level PPO into token-level problems and derives a framework for token-level reward guidance in DPO, enabling more fine-grained policy optimization.

Findings

Achieves up to 7.5% win rate improvement on MT-Bench

Demonstrates significant performance gains over standard DPO

Provides a practical reward guidance framework for token-level optimization

Abstract

Recent advancements in reinforcement learning from human feedback have shown that utilizing fine-grained token-level reward models can substantially enhance the performance of Proximal Policy Optimization (PPO) in aligning large language models. However, it is challenging to leverage such token-level reward as guidance for Direct Preference Optimization (DPO), since DPO is formulated as a sequence-level bandit problem. To address this challenge, this work decomposes the sequence-level PPO into a sequence of token-level proximal policy optimization problems and then frames the problem of token-level PPO with token-level reward guidance, from which closed-form optimal token-level policy and the corresponding token-level reward can be derived. Using the obtained reward and Bradley-Terry model, this work establishes a framework of computable loss functions with token-level reward guidance for DPO, and proposes a practical reward guidance based on the induced DPO reward. This formulation enables different tokens to exhibit varying degrees of deviation from reference policy based on their respective rewards. Experiment results demonstrate that our method achieves substantial performance improvements over DPO, with win rate gains of up to 7.5 points on MT-Bench, 6.2 points on AlpacaEval 2, and 4.3 points on Arena-Hard. Code is available at https://github.com/dvlab-research/TGDPO.