Provably Mitigating Corruption, Overoptimization, and Verbosity Simultaneously in Offline and Online RLHF/DPO Alignment
Ziyi Chen, Junyi Li, Peiran Yu, Heng Huang
TL;DR
This work tackles three intertwined problems in RLHF/DPO-based LLM alignment—data corruption, reward overoptimization, and verbosity. It introduces RLHF-COV and DPO-COV to jointly address these issues in offline and online regimes by combining noise modeling, pessimistic/optimistic regularization, and length penalties, and proves equivalence with vanilla RLHF/DPO in the reward-induced policy space. The offline analysis yields length-regularized generalization bounds that extend prior results to corrupted data and verbosity control, while the online analysis provides corresponding guarantees. Empirical results on Argilla and math/reasoning tasks demonstrate improved length-controlled performance and robustness, supporting practical applicability without additional reward-model estimation overhead.
Abstract
Reinforcement learning from human feedback (RLHF) and direct preference optimization (DPO) are important techniques to align large language models (LLM) with human preference. However, the quality of RLHF and DPO training is seriously compromised by \textit{\textbf{C}orrupted} preference, reward \textit{\textbf{O}veroptimization}, and bias towards \textit{\textbf{V}erbosity}. To our knowledge, most existing works tackle only one of these important issues, and the few other works require much computation to estimate multiple reward models and lack theoretical guarantee of generalization ability. In this work, we propose RLHF-\textbf{COV} and DPO-\textbf{COV} algorithms that can simultaneously mitigate these three issues, in both offline and online settings. This ability is theoretically demonstrated by obtaining length-regularized generalization error rates for our DPO-COV algorithms trained on corrupted data, which match the best-known rates for simpler cases with clean data and without length regularization. Moreover, our DPO-COV algorithm is simple to implement without reward estimation, and is proved to be equivalent to our RLHF-COV algorithm, which directly implies the equivalence between the vanilla RLHF and DPO algorithms. Experiments demonstrate the effectiveness of our DPO-COV algorithms under both offline and online settings.