Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning

2026-07-08

AI Summary

This work addresses the training instability in asynchronous reinforcement learning for long-horizon agent tasks, which stems from off-policy bias and mismatched batch sampling. To mitigate these issues, the authors propose the Single-trajectory Asynchronous Optimization (SAO) framework, which replaces conventional batch sampling with single-trajectory sampling per prompt to reduce off-policy bias. SAO further incorporates a bilateral token-level clipping mechanism jointly trained with a value model, significantly enhancing optimization stability and generalization. Experimental results demonstrate that SAO consistently outperforms GRPO and its variants across multiple benchmarks, including SWE-Bench Verified, BeyondAIME, and IMOAnswerBench, and has been successfully deployed in the reinforcement learning training of the GLM-5.2 (750B-A40B) agent model.

Abstract

Reinforcement learning (RL) is becoming increasingly important for post-training large language models (LLMs). Previous RL pipelines for LLMs were mostly synchronous and batch-interleaved, which is inefficient for long-horizon agentic tasks. Recently, asynchronous RL has emerged as a more efficient alternative by updating the model as rollouts arrive. However, existing asynchronous RL systems often emphasize throughput, while leaving training stability and task effectiveness largely underexplored. For example, a key challenge is that group-wise sampling in the widely-used GRPO framework does not naturally fit asynchronous agentic training. In this paper, we present Single-rollout Asynchronous Optimization (SAO) to address the stability and off-policy challenges in asynchronous RL. To reduce off-policy effects and improve generalization, we replace group-wise sampling with single-rollout sampling, that is, using one rollout per prompt. We further improve this single-rollout strategy with practical value-model training designs. To improve optimization stability, we introduce a strict double-side token-level clipping strategy. SAO is able to train stably for one thousand steps and consistently outperform GRPO and its variants on agentic coding and reasoning benchmarks, such as SWE-Bench Verified, BeyondAIME, and IMOAnswerBench. We also demonstrate that single-rollout RL is particularly effective in a simulated online learning setting, where the model must adapt to changing evolving environments. To this end, SAO is successfully deployed in the agentic RL pipeline for training the open GLM-5.2 model (750B-A40B).

Problem

• asynchronous reinforcement learning
• training stability
• off-policy effects
• agentic tasks
• large language models

Innovation

• Single-rollout Sampling
• Asynchronous Reinforcement Learning
• Token-level Clipping
• Off-policy Mitigation
• Agentic RL