arXiv:2605.15466v2 Announce Type: replace 
Abstract: Learning predictive world models from unlabelled video is a foundational challenge in artificial intelligence. While Joint Embedding Predictive Architectures (JEPA) have set new benchmarks in semantic classification, they often remain physics-blind, failing to capture the causal dynamics necessary for downstream reasoning. We hypothesize that this stems from standard patch-based masking strategies, which prioritize visual texture over rare but informative kinematic events. We propose Interaction-Aware JEPA (IA-JEPA), which utilizes a self-supervised motion-centric masking strategy to prioritize physical interactions. By specifically targeting entities engaged in collisions or momentum transfers, we force the architecture to reconstruct latent trajectories rather than static background features. Evaluated on the CLEVRER benchmark, IA-JEPA achieves 14.26% accuracy on causal reasoning tasks, a significant lead over the 3.22% achieved by standard patch-masked baselines. Crucially, we demonstrate that IA-JEPA breaks the "static bias" of standard self-supervision by inducing a higher-entropy, more discriminative latent space (+10% entropy gain) that linearizes physical energy ($R^2=0.43$). We show that this interaction bias generalizes to real-world human actions (Something-Something V2) and zero-shot physical puzzles (PHYRE-Lite). Our results provide a scalable, fully self-supervised path toward building foundational world models that begin to internalize the causal structure of the physical world.

أدت التطورات الأخيرة في الذكاء الاصطناعي إلى تطوير نماذج تنبؤية تعتمد على التضمين المشترك (IA-JEPA)، والتي تعزز من توقعات الفيديو السببي من خلال التركيز على التفاعلات الفيزيائية بدلاً من الميزات الثابتة. وقد أظهرت هذه الطريقة تحسينات كبيرة في الدقة على معيار CLEVRER، حيث حققت 14.26% في مهام التفكير السببي.

Los recientes avances en inteligencia artificial han llevado al desarrollo de arquitecturas predictivas de incrustación conjunta (IA-JEPA), que mejoran la predicción de video causal al centrarse en las interacciones físicas en lugar de características estáticas. Este enfoque ha mostrado mejoras significativas en precisión en el benchmark CLEVRER, alcanzando un 14.26% en tareas de razonamiento causal.

Des avancées récentes en intelligence artificielle ont conduit au développement d'architectures prédictives d'embedding joint (IA-JEPA), qui améliorent la prédiction vidéo causale en se concentrant sur les interactions physiques plutôt que sur des caractéristiques statiques. Cette approche a montré des améliorations significatives de précision sur le benchmark CLEVRER, atteignant 14,26 % dans les tâches de raisonnement causal.

Recent advancements in artificial intelligence have led to the development of Interaction-Aware Joint Embedding Predictive Architectures (IA-JEPA), which enhances causal video prediction by focusing on physical interactions rather than static features. This approach has shown significant improvements in accuracy on the CLEVRER benchmark, achieving 14.26% in causal reasoning tasks.