OneWM-VLA: Compressing Visual Stream to Single Token per Frame

Researchers propose OneWM-VLA, a method that compresses each video frame into a single semantic token for world-model-augmented vision-language-action (VLA) models. Existing approaches pass high-bandwidth visual streams into world modules, leaving per-frame representation and action coupling under-examined under constrained adaptation budgets. OneWM-VLA uses Adaptive Attention Pooling to reduce per-frame visual bandwidth to one token, and produces latent stream and action trajectory under a single flow-matching objective. Empirical results show no compromise in performance.