Shuhan Tan

Driving world models from ~100M to 1.3B params: frozen DINOv3 S/B/L encoder + trained predictor, with and without action expert; plus a 1B reference.