Asynchronous RLHF: Faster and More Efficient Off-Policy RL for Language Models

TL;DR

This paper introduces asynchronous RLHF, a method that separates generation and learning to enable faster, off-policy training of language models, achieving significant speedups while maintaining performance.

Contribution

It proposes a novel asynchronous RLHF framework that improves training efficiency and scalability for language models by decoupling sample generation from learning.

Findings

Asynchronous RLHF speeds up training by ~40% for chatbots.

Robustness to off-policy data increases with model size.

Achieves ~70% faster fine-tuning on GSM8k with maintained accuracy.

Abstract

The dominant paradigm for RLHF is online and on-policy RL: synchronously generating from the large language model (LLM) policy, labelling with a reward model, and learning using feedback on the LLM's own outputs. While performant, this paradigm is computationally inefficient. Inspired by classical deep RL literature, we propose separating generation and learning in RLHF. This enables asynchronous generation of new samples while simultaneously training on old samples, leading to faster training and more compute-optimal scaling. However, asynchronous training relies on an underexplored regime, online but off-policy RLHF: learning on samples from previous iterations of our model which give a worse training signal. We tackle the fundamental challenge in this regime: how much off-policyness can we tolerate for asynchronous training to speed up learning but maintain performance? Among several RLHF algorithms we test, online DPO is found to be most robust to off-policy data, and robustness increases with the scale of the policy model. We study further compute optimizations for asynchronous RLHF but find that they come at a performance cost, giving rise to a trade-off. We verify the scalability of asynchronous RLHF by training a general-purpose chatbot from LLaMA 3.1 8B on an instruction-following task ~40% faster than a synchronous run while matching final performance. Finally, we extend our results to math and reasoning to demonstrate asynchronous RL can finetune Rho 1B on GSM8k ~70% faster while matching synchronous accuracy.