Dual-Plane World Model v2: Attention Directionality in Physics-Language Composition

We extend the dual-plane world model (v1, March 2026) with experiments on real video data (UCF-101). We introduce Multi-Cross Attention and show that physics-as-query attention outperforms symmetric attention. Physical representations remain recoverable without linguistic context (32.2% degradation) and are robust to semantic corruption (+18.7%). Code: github.com/Artyis/dpwm Keywords: world models, JEPA, multimodal learning, physical reasoning, attention mechanism, self-supervised learning, reversible composition, video understanding, UCF-101, LeCun