OPSD-V introduces an on-policy self-distillation framework to enhance post-training few-step autoregressive video diffusion models, addressing error accumulation and degraded motion dynamics in long video generation. By incorporating real long-video data as temporal context during training, the method mitigates long-horizon performance drops while maintaining the original few-step inference efficiency. The approach aims to improve the quality and consistency of generated videos in autoregressive rollout scenarios.