Rethinking Interaction Modeling in World Models

Traditional world models often struggle to scale when representing multi-agent or complex object interactions because they rely on monolithic state representations. The IMPACT (Interaction Map for Scalable Interaction-Aware World Model Training) framework addresses this by treating attention mechanisms not just as a feature-weighting tool, but as an explicit, learnable interaction map. By doing so, the model gains the ability to decouple individual agent states from their relational dynamics, allowing for more robust predictions in dynamic environments.

The Mechanism: Attention as a Relational Map

The core innovation of IMPACT is the integration of an attention-based interaction map that dynamically identifies which entities in a scene are influencing one another. Instead of forcing the model to learn global dependencies, the attention map acts as a sparse, interpretable structure that guides the world model's predictive head. This approach reduces the computational overhead typically associated with modeling high-dimensional interactions, enabling the system to scale to environments with a higher number of entities without a linear increase in complexity. By explicitly mapping interactions, the model achieves better generalization in scenarios where agents or objects enter or leave the scene, as the attention map naturally adapts to the current set of active participants.