Zero Mismatch, Simple Async RL: REINFORCE Zero, DeepSeek-V4.1 and Kimi K3.1

The REINFORCE Zero paper, from NVIDIA Research and Inferact (Jian Hu et al.), argues that asynchronous reinforcement learning can be stabilized by aligning rollout and training kernels, so that the remaining gap from policy lag is handled by plain unclipped importance sampling.