Provably Efficient Online RLHF with One-Pass Reward Modeling

Abstract: Reinforcement Learning from Human Feedback (RLHF) has shown remarkable success in aligning LLMs with human preferences. Traditional RLHF approaches rely on a fixed dataset, which often suffers from limited coverage. To this end, online RLHF has emerged as a promising direction, enabling iterative data collection and model improvement. Despite its potential, this paradigm faces a key bottleneck: the requirement to continuously integrate new data into the historical dataset and re-optimize the model from scratch at each iteration, resulting in computational and storage costs that grow linearly with the number of iterations. In this work, we address this challenge by proposing a one-pass reward modeling method that does not require storing the historical data and can be computed in constant time. Specifically, we first formalize RLHF as a contextual preference bandit problem and design an online mirror descent algorithm with a tailored local norm to replace the standard maximum likelihood estimation for reward modeling. We then apply our method to various online RLHF settings, including passive data collection, active data collection, and deployment-time adaptation. We provide theoretical guarantees showing that our method improves both statistical and computational efficiency. Finally, we provide practical algorithms and conduct experiments using Llama-3-8B-Instruct and Qwen2.5-7B-Instruct models on the Ultrafeedback-binarized and Mixture2 datasets, validating the effectiveness of our proposed method.