IMPACT Improves World Models for Physically Plausible Interactions.

World models have shown significant progress in predicting future actions for embodied agents, but they often struggle to accurately model physically plausible interactions. Existing solutions typically rely on external representations like motion or geometry, which are costly to obtain. This research identifies a core problem: the globally averaged mean squared error (MSE) denoising objective disproportionately emphasizes static content, undersupervising the sparse, dynamic regions critical for interaction generation. To overcome this, the IMPACT framework (Interaction-aware Model training with Prior-guided Attention Calibration and Targeting) is introduced. IMPACT leverages cross-attention associated with manipulated-object tokens as an internal prior for action-conditioned changes. It then samples candidate regions from this prior, calibrates them with local prediction errors, and constructs an "interaction map" to reweight denoising supervision. This approach significantly improves interaction fidelity, physical plausibility, and visual quality across various robot-arm and human-hand manipulation tasks, without requiring external representations or inference-time modifications.