中文HorizonDrive：面向长时域驾驶仿真的自校正自回归世界模型

闭环驾驶模拟需要自回归（AR）滚动以实现实时交互。现有AR蒸馏方法存在缺陷：帧下沉方式因自车快速运动和场景快速变化而迁移效果差；学生端退化训练受限于教师的单次输出长度，监督视野有限。关键问题在于：能否通过AR扩展教师模型本身？这提示了现有方法在长时程动态场景下的局限性，亟需更有效的AR扩展策略来提升闭环仿真性能。

arXiv:2605.11596v2 Announce Type: replace Abstract: Closed-loop driving simulation requires real-time interaction beyond short offline clips, pushing current driving world models toward autoregressive (AR) rollout. Existing AR distillation approaches typically rely on frame sinks or student-side degradation training. The former transfers poorly to driving due to fast ego-motion and rapid scene changes, while the latter remains bounded by the teacher's single-pass output length and thus provides only a limited supervision horizon. A natural question is: can the teacher itself be extended via AR