Self-Play with Adversarial Critic: Provable and Scalable Offline Alignment for Language Models

TL;DR

This paper introduces SPAC, a scalable and theoretically guaranteed offline preference optimization method with self-play for aligning large language models, demonstrating both convergence proofs and competitive empirical results.

Contribution

It presents SPAC, the first provable and scalable offline alignment method for LLMs, combining theoretical guarantees with practical effectiveness.

Findings

SPAC converges under single-policy concentrability.

SPAC performs competitively on a 7B Mistral model.

Theoretical analysis supports its effectiveness in large-scale settings.

Abstract

This work studies the challenge of aligning large language models (LLMs) with offline preference data. We focus on alignment by Reinforcement Learning from Human Feedback (RLHF) in particular. While popular preference optimization methods exhibit good empirical performance in practice, they are not theoretically guaranteed to converge to the optimal policy and can provably fail when the data coverage is sparse by classical offline reinforcement learning (RL) results. On the other hand, a recent line of work has focused on theoretically motivated preference optimization methods with provable guarantees, but these are not computationally efficient for large-scale applications like LLM alignment. To bridge this gap, we propose SPAC, a new offline preference optimization method with self-play, inspired by the on-average pessimism technique from the offline RL literature, to be the first provable and scalable approach to LLM alignment. We both provide theoretical analysis for its convergence under single-policy concentrability for the general function approximation setting and demonstrate its competitive empirical performance for LLM alignment on a 7B Mistral model with Open LLM Leaderboard evaluations.