The work decomposes the rollout-to-learner gap in asynchronous RL into smooth weight lag, which unclipped importance sampling can correct, and fluctuating kernel mismatch, which it cannot. Aligning kernels enables stable training for 20,000 updates at policy lag 64 and 2.40x the throughput of synchronous training, with support for DeepSeek-V4.1-Flash, Kimi K3.1, and AMD MI355X.
Asynchronous RL
异步强化学习
An RL paradigm where trainer and inference engine work in parallel without strict synchronous waiting.
Curated Intelligence
Hugging Face's TRL library introduces 'Delta Weight Sync', which transmits only model updates instead of full weights, drastically reducing bandwidth requirements in Asynchronous Reinforcement Learning (Async RL). This optimization addresses the bottleneck of transferring terabytes of data per step, making RL fine-tuning for trillion-parameter models feasible on existing infrastructure.
Zhipu AI has officially launched GLM-5, a next-generation flagship model designed to shift coding paradigms from 'Vibe Coding' to 'Agentic Engineering'. By integrating DeepSeek Sparse Attention (DSA) for cost efficiency and a novel asynchronous reinforcement learning infrastructure, the model achieves state-of-the-art performance in real-world software engineering tasks, rivaling Claude Opus 4.5 and GPT-5.2.
Zhipu AI has open-sourced GLM-5, scaling parameters from 355B to 744B (40B active) and pre-training data to 28.5T tokens. The model achieves open-source SOTA on benchmarks like SWE-bench-Verified, with coding and agent capabilities approaching Claude Opus 4.5. Key innovations include the 'Slime' asynchronous reinforcement learning framework and DeepSeek Sparse Attention, enabling efficient handling of complex engineering tasks and long-horizon agentic workflows.