Modality-Autoregressive World-Action Models

The paper presents ModAR, the first world-action model (WAM) to autoregressively denoise multiple future modalities before predicting actions, enabling each prediction to condition on previously generated modalities. Evaluations show that ModAR outperforms existing WAM formulations, achieving a higher average success rate with fewer training FLOPs and no pretraining, particularly in real-world bimanual tasks.