Multi-turn Reinforcement Learning from Preference Human Feedback

TL;DR

This paper introduces novel multi-turn preference-based reinforcement learning methods for aligning large language models, demonstrating improved performance over existing single-turn approaches in dialogue and education environments.

Contribution

It develops a new mirror-descent-based policy optimization algorithm for multi-turn preference RL and proves its convergence, addressing limitations of single-turn preference methods.

Findings

Deep RL variant outperforms RLHF baselines in Education Dialogue environment.

Algorithm recovers reward-based RL performance using only preference signals.

Proven convergence to Nash equilibrium in the tabular setting.

Abstract

Reinforcement Learning from Human Feedback (RLHF) has become the standard approach for aligning Large Language Models (LLMs) with human preferences, allowing LLMs to demonstrate remarkable abilities in various tasks. Existing methods work by emulating the preferences at the single decision (turn) level, limiting their capabilities in settings that require planning or multi-turn interactions to achieve a long-term goal. In this paper, we address this issue by developing novel methods for Reinforcement Learning (RL) from preference feedback between two full multi-turn conversations. In the tabular setting, we present a novel mirror-descent-based policy optimization algorithm for the general multi-turn preference-based RL problem, and prove its convergence to Nash equilibrium. To evaluate performance, we create a new environment, Education Dialogue, where a teacher agent guides a student in learning a random topic, and show that a deep RL variant of our algorithm outperforms RLHF baselines. Finally, we show that in an environment with explicit rewards, our algorithm recovers the same performance as a reward-based RL baseline, despite relying solely on a weaker preference signal.