Adnan KhanAdnanProduct & AI
San Francisco Bay Area
Get in touch

Source note3 min read

Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning

Source

arXiv

arxiv.org/abs/2607.07508v1

Read the source

Source

Plain-Language Brief

The paper proposes a new method called Single-rollout Asynchronous Optimization (SAO) for Reinforcement Learning (RL) in post-training large language models (LLMs). SAO aims to improve stability and task effectiveness, addressing the limitations of previous asynchronous RL systems. The method uses single-rollout sampling instead of group-wise sampling, enhances it with practical value-model training designs, and introduces a strict double-side token-level clipping strategy for optimization stability. SAO consistently outperforms other methods on various agentic coding and reasoning benchmarks and is effective in simulated online learning settings.

What the Document Says

  • The paper discusses the application of Reinforcement Learning (RL) for post-training large language models (LLMs).
  • Previous RL pipelines for LLMs were mostly synchronous and batch-interleaved, but are inefficient for long-horizon agentic tasks.
  • Asynchronous RL has emerged as a more efficient alternative by updating the model as rollouts arrive, but existing systems often emphasize throughput while leaving training stability and task effectiveness underexplored.
  • A key challenge is that group-wise sampling in GRPO framework doesn't fit well with asynchronous agentic training.
  • The authors present Single-rollout Asynchronous Optimization (SAO) to address these issues.
  • SAO uses single-rollout sampling for reducing off-policy effects and improving generalization.
  • It also introduces a strict double-side token-level clipping strategy for optimization stability.
  • SAO consistently outperforms GRPO and its variants on agentic coding and reasoning benchmarks such as SWE-Bench Verified, BeyondAIME, and IMOAnswerBench.
  • SAO is effective in simulated online learning settings.
  • SAO has been successfully deployed in the agentic RL pipeline for training the open GLM-5.2 model (750B-A40B).

Key Concepts Mentioned

  • Reinforcement Learning
  • Large Language Models
  • Asynchronous Reinforcement Learning
  • Single-rollout Sampling
  • Double-side Token-Level Clipping
  • Agentic Coding and Reasoning Benchmarks
  • Simulated Online Learning
  • GLM-5.2 Model

Claims Made by the Document

  • SAO improves stability and task effectiveness in asynchronous RL.
  • Single-rollout sampling reduces off-policy effects and improves generalization.
  • Double-side token-level clipping improves optimization stability.

Evidence Used

  • The paper presents SAO as a solution to stability and off-policy challenges in asynchronous RL.
  • SAO uses single-rollout sampling instead of group-wise sampling for reducing off-policy effects.
  • SAO introduces a strict double-side token-level clipping strategy for optimization stability.

Methods / Approach

UNKNOWN

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

UNKNOWN

Open Questions

UNKNOWN

Notes for Future Q&A

May include interpretation. Labeled where it does.

UNKNOWN

Raw Abstract / Excerpt

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).