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

2025-06-17

AI Summary

Direct Preference Optimization (DPO) struggles to leverage fine-grained, token-level reward signals because it inherently models preference learning as a sequence-level bandit problem. Method: This paper proposes Token-level DPO (T-DPO), which decomposes sequence-level policy optimization into token-level subproblems. Grounded in the Bradley–Terry model, T-DPO derives closed-form optimal token-level policies and rewards, and introduces a reward-guided token-level DPO loss function. Contribution/Results: T-DPO is the first framework to explicitly incorporate token-level reward signals into DPO training, enabling each token to dynamically adjust its deviation from the reference policy based on local reward estimates—thus overcoming DPO’s inherent sequence-level modeling constraint. Empirical evaluation shows substantial improvements over standard DPO: +7.5, +6.2, and +4.3 win-rate points on MT-Bench, AlpacaEval 2, and Arena-Hard, respectively.

Technology Category

• Search and Optimization: Learning to Search
• Machine Learning: Learning Preferences or Rankings
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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.

Problem

• Enhance DPO with token-level reward guidance
• Decompose sequence-level PPO into token-level problems
• Improve alignment performance using fine-grained rewards

Innovation

• Token-level reward guidance for DPO enhancement
• Decomposes sequence-level PPO into token-level problems
• Computable loss functions with token-level rewards