Abstract
Reinforcement Learning with Human Feedback (RLHF) has become crucial for aligning Large Language Models (LLMs) with human intent. However, existing offline RLHF approaches suffer from overoptimization, where language models degrade by overfitting inaccuracies and drifting from preferred behaviors observed during training. Distributionally robust optimization (DRO) is a natural solution, but existing DRO-DPO methods are sample-inefficient, ignore heterogeneous preferences, and lean on brittle heuristics. We introduce \emph{DRO-REBEL}, a family of robust online REBEL updates built on type-p Wasserstein, Kullback-Leibler (KL), and \chi^2 ambiguity sets. Strong duality reduces each update to a relative-reward regression, retaining REBEL's scalability without PPO-style clipping or value networks. Under linear rewards, log-linear policies, and a standard coverage condition, we prove \widetilde{O}(\sqrt{d/n}) bounds on squared parameter error, with sharper constants than prior DRO-DPO analyses, and give the first parametric \widetilde{O}(d/n) rate for DRO-based alignment under preference shift, matching non-robust RLHF in benign regimes. Each divergence yields a tractable SGD-based algorithm: gradient regularization for Wasserstein, importance weighting for KL, and a 1-D dual solve for \chi^2. On Emotion Alignment, the ArmoRM multi-objective benchmark, and HH-Alignment, DRO-REBEL outperforms prior robust and non-robust baselines across unseen preference mixtures, model sizes, and dataset scales.
