AReaL v0.3: Open-Source Async RL Training System Now Supports Agentic Model Training

inclusionAI / GitHub·medium signal

AReaL (developed by Tsinghua IIIS and Ant Group) released v0.3 adding support for agentic model training on top of its existing math reasoning and coding RL pipelines. The system fully decouples rollout workers from training workers — rollout runs continuously while gradients are applied asynchronously — achieving up to 2.77x speedup over synchronous RL baselines with higher GPU utilization. Open-source at github.com/inclusionAI/AReaL, directly relevant for anyone training reasoning or agent models.