Kairos01：面向 Physical AI 的具备后悔感知能力的原生世界-动作模型栈【注意力：①滑动窗口注意力SWA；②扩张滑动窗口注意力DSWA；③门控线性全局注意力GLA】

Kairos: A Regret-Aware Native World-Action Model Stack for Physical AI

Kairos：面向 Physical AI 的具备后悔感知能力的原生世界-动作模型栈

Learning, Maintaining, and Deploying Control-Sufficient World States Kairos Team

Abstract 摘要

World models are becoming a central substrate for Physical AI, yet their current development is differentiating across representational, generative, interactive, and unified understandinggenerationprediction world models.

While these paradigms have advanced abstract world reasoning, high-fidelity visual generation, interactive simulation, and embodied world-action modeling, none alone provides the deployable, long-horizon, control-sufficient state required by embodied agents. We intr