ReWAM: Reciprocal World Action Models for Interactive Autonomous Driving
This study addresses the limited action generation performance of existing models in dense interactive scenarios, where surrounding vehicles are not treated as decision-makers. To this end, we propose a game-theoretic reciprocal world-action modeling framework that captures the mutual influence between the ego vehicle and surrounding agents via Level-k reasoning hierarchies. By integrating shared world representations, a Diffusion Transformer (DiT) architecture, and cross-agent attention mechanisms, the framework achieves interaction-aware action generation. Notably, this work pioneers treating surrounding vehicles as conditional responders and learns expert policies through joint optimization across the entire reasoning hierarchy. Evaluations on the NAVSIM dataset demonstrate state-of-the-art performance, with particularly significant improvements in highly interactive scenarios, thereby validating the effectiveness of reciprocal response modeling.