Policy Gradient Primal-Dual Method for Safe Reinforcement Learning from Human Feedback
Provides the first theoretical convergence guarantees for safe RLHF under an infinite horizon CMDP, addressing a key gap in prior empirical work.
This paper formulates safe RLHF as an infinite horizon discounted CMDP and proposes two primal-dual algorithms that avoid reward model fitting and support flexible trajectory lengths, achieving global convergence guarantees with polynomial rates.
Safe Reinforcement Learning from Human Feedback (Safe RLHF) has recently achieved empirical success in developing helpful and harmless large language models by decoupling human preferences regarding helpfulness and harmlessness. Existing approaches typically rely on fitting fixed horizon reward models from human feedback and have only been validated empirically. In this paper, we formulate safe RLHF as an infinite horizon discounted Con- strained Markov Decision Process (CMDP), since humans may interact with the model over a continuing sequence of interactions rather than within a single finite episode. We propose two Safe RLHF algorithms that do not require reward model fitting and, in contrast to prior work assuming fixed-length trajectories, support flexible trajectory lengths for training. Both algo- rithms are based on the primal-dual method and achieve global convergence guarantees with polynomial rates in terms of policy gradient iterations, trajectory sample lengths, and human preference queries. To the best of our knowledge, this is the first work to study infinite horizon discounted CMDP under human feedback and establish global, non-asymptotic convergence.