REBEL: Reinforcement Learning via Regressing Relative Rewards

TL;DR

REBEL is a minimalist reinforcement learning algorithm that simplifies policy optimization by regressing relative rewards, matching or surpassing PPO's performance while being easier to implement and more efficient.

Contribution

It introduces REBEL, a lightweight RL algorithm that reduces policy optimization to relative reward regression, with strong theoretical guarantees and practical success in language and image tasks.

Findings

REBEL matches or exceeds PPO and DPO performance.

REBEL is simpler to implement and more computationally efficient.

REBEL performs well in fine-tuning large language models.

Abstract

While originally developed for continuous control problems, Proximal Policy Optimization (PPO) has emerged as the work-horse of a variety of reinforcement learning (RL) applications, including the fine-tuning of generative models. Unfortunately, PPO requires multiple heuristics to enable stable convergence (e.g. value networks, clipping), and is notorious for its sensitivity to the precise implementation of these components. In response, we take a step back and ask what a minimalist RL algorithm for the era of generative models would look like. We propose REBEL, an algorithm that cleanly reduces the problem of policy optimization to regressing the relative reward between two completions to a prompt in terms of the policy, enabling strikingly lightweight implementation. In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL, which allows us to match the strongest known theoretical guarantees in terms of convergence and sample complexity in the RL literature. REBEL can also cleanly incorporate offline data and be extended to handle the intransitive preferences we frequently see in practice. Empirically, we find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO, all while being simpler to implement and more computationally efficient than PPO. When fine-tuning Llama-3-8B-Instruct, REBEL achieves strong performance in AlpacaEval 2.0, MT-Bench, and Open LLM Leaderboard.