NVIDIA and Inferact release REINFORCE Zero for async RL

Researchers have introduced REINFORCE Zero, a recipe utilizing aligned kernels and unclipped importance sampling to stabilize asynchronous reinforcement learning. By removing kernel mismatch, the method allows models to train stably at policy lags up to 64 while achieving significantly higher training throughput. The approach has been validated across flagship models including DeepSeek-V4.1-Flash and Kimi K3.1 on both NVIDIA and AMD hardware. Open article →