$ cat zero-mismatch-simple-async-rl-reinforce-zero-deepseek-v41-and-kimi-k31.md
Zero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1
Explains how asynchronous reinforcement learning can train on stale model weights, and how aligning rollout and training kernels bit for bit, together with unclipped importance sampling, corrects that lag. Covers zero-mismatch RL applied to DeepSeek-V4.1-Flash, Kimi K3.1 and AMD MI355X hardware.
vLLM06 Nov 2026 · 1 min read · summary
Read on vllm.ai ↗Asynchronous RL trains on stale weights. Align rollout and training kernels bit for bit, and unclipped importance sampling corrects that lag. Plus zero-mismatch RL for DeepSeek-V4.1-Flash, Kimi K3.1 and AMD MI355X.
source
This is a summary of an article first published by vLLM. The full piece is on their site.