Researchers present ReWorld, an interactive world model that addresses the competing demands of real-time control and long-horizon memory through architectural separation, using mixed per-head attention windows where most heads focus on recent history while global heads span entire sequences, complemented by a pose-indexed landmark bank for efficient retrieval during inference. Evaluated against six comparable models, ReWorld shows superior control precision with 11.95-degree rotation error and strong video quality while maintaining a fixed 12-chunk cache that successfully regenerates starting views across minute-long trajectories.
