GAC: Stabilizing Asynchronous RL Training for LLMs via Gradient Alignment Control

TL;DR

This paper introduces GAC, a method to stabilize asynchronous reinforcement learning training of large language models by controlling gradient alignment, addressing instability caused by stale, correlated gradients.

Contribution

The paper identifies the cause of instability in asynchronous RL as high cosine similarity between consecutive gradients and proposes GAC to mitigate this issue with theoretical guarantees.

Findings

GAC stabilizes asynchronous RL training at high staleness levels.

Asynchronous training exhibits persistently high cosine similarity between gradients.

GAC matches synchronized training performance despite high asynchrony.

Abstract

Asynchronous execution is essential for scaling reinforcement learning (RL) to modern large model workloads, including large language models and AI agents, but it can fundamentally alter RL optimization behavior. While prior work on asynchronous RL focuses on training throughput and distributional correction, we show that naively applying asynchrony to policy-gradient updates can induce qualitatively different training dynamics and lead to severe training instability. Through systematic empirical and theoretical analysis, we identify a key signature of this instability: asynchronous training exhibits persistently high cosine similarity between consecutive policy gradients, in contrast to the near-orthogonal updates observed under synchronized training. This stale-aligned gradient effect amplifies correlated updates and increases the risk of overshooting and divergence. Motivated by this observation, we propose GRADIENT ALIGNMENT CONTROL(GAC), a simple dynamics-aware stabilization method that regulates asynchronous RL progress along stale-aligned directions via gradient projection. We establish convergence guarantees under bounded staleness and demonstrate empirically that GAC recovers stable, on-policy training dynamics and matches synchronized baselines even at high staleness.