Z.AI'S STABLE ASYNCHRONOUS RL

Summary from the TLDR AI newsletter

Single-rollout Asynchronous Optimization replaces GRPO-style grouped sampling with one rollout per prompt, supported by value-model training and strict token-level clipping.