Zhipu's SAO: Stable Asynchronous RL for Long-Horizon Agents

Zhipu released SAO, a single-rollout asynchronous optimization method that stabilizes RL training for long-horizon agents over a thousand steps, outperforming GRPO and suggesting new scaling directions beyond parameter size.

2026-08-21 ~ 2026-08-21 · 2 related posts

• Zhipu's SAO: single-rollout async RL trains stably for 1,000 steps, beats GRPO — teortaxesTex · 2026-08-21
• Paper: SAO explores async optimization for long-horizon agents — khademinori · 2026-08-21