Tsinghua's SAO Ditches GRPO Groups, Trains Open GLM-5.2

A team from Tsinghua University has taken a swing at one of the more annoying problems in current LLM post-training: asynchronous reinforcement learning that collapses before you can run it long enough to matter. Their fix, described in a paper hosted on Hugging Face, is called Single-rollout Asynchronous Optimization, or SAO, and it argues that the group-based sampling underlying GRPO is exactly what breaks asynchronous agentic training.

The claim comes with a specific comparison. Vanilla GRPO, the authors report, suffers a performance collapse at approximately 160 training steps in their asynchronous setup; SAO trains stably for around one thousand steps on the same benchmarks. They get there by replacing group-wise sampling with one rollout per prompt, adding a strict double-sided token-level clipping strategy the authors call DIS, and re-engineering the value model with faster critic updates and a frozen-attention scheme in which the attention parameters are held fixed while the MoE projections train. The gains show up on SWE-Bench Verified, where SAO reaches 29.8% against 27.0% for GRPO patched with the same clipping strategy and 23.0% for the raw Qwen3-30B-A3B baseline. On math under a Python-tool setting, SAO scores 97.3% on AIME2025 and 74.8% on BeyondAIME.

The reason to care is where the pipeline ends up. The authors say SAO was deployed in the agentic RL pipeline used to train the open GLM-5.2 model, which they describe as 750B-A40B. If you are an open-model shop trying to catch up to closed frontier systems on long-horizon agentic tasks, the recipe (drop GRPO's synchronization barrier, invest in a real critic instead of pretending you do not need one) is now sitting in the open.

The honest caveats are the usual ones. The results are single-lab numbers, the ablations run only on the Qwen3-30B-A3B backbone, and the paper does not give wall-clock or GPU-hour costs for the GLM-5.2 run itself. Whether the collapse-at-160-steps pattern holds in other labs' async setups or is somewhat specific to this configuration is the kind of thing only replication settles. Take the specifics as reported, not as settled.

Still, the direction is the interesting bit. For a couple of years the open-agent field has been optimizing GRPO variants for stability. If a Tsinghua team is now publicly saying group-sampling is the wrong primitive for asynchronous agentic training, and using the alternative to train an open large-scale model, other labs building long-rollout coding and reasoning agents have a fresh design question to answer.