CST-WM: A Causally Structured World Model for Embodied Visual Tracking
The episode discusses the paper "CST-WM: A Causally Structured World Model for Embodied Visual Tracking" from NYU Abu Dhabi. The hosts explain how this model decomposes the latent state into distinct branches to enforce causal constraints, improving long-horizon tracking and target re-acquisition by separating action effects from target evidence updates.