English

One Token Per Frame: Reconsidering Visual Bandwidth in World Models for VLA Policy

Computer Vision and Pattern Recognition 2026-05-15 v3 Artificial Intelligence

Abstract

Vision-language-action (VLA) models increasingly rely on auxiliary world modules to plan over long horizons, yet how such modules should be parameterized on top of a pretrained VLA remains an open design question. Existing world-model-augmented VLAs typically pass the per-frame visual stream into the world module at high visual bandwidth and treat its rollout as a side product of action prediction; under a constrained adaptation budget on a frozen backbone, this leaves both the per-frame representation and the latent action coupling under-examined. We introduce OneWM-VLA, which compresses each view into a single semantic token per frame through an Adaptive Attention Pooling, and produces the resulting latent stream and the action trajectory under a single flow-matching objective rather than connecting them through a separate decoder. Empirically, we find that per-frame visual bandwidth can be reduced to a single token without compromising long-horizon performance under our setup. Trained with 14.71M LoRA parameters on a π0\pi_0 (2B) backbone, OneWM-VLA improves the average success rate from 47.9% to 61.3% on MetaWorld~MT50, reaches 95.6% on LIBERO-Long (vs.85.2% for π0\pi_0), and reaches 60.0% on the long-horizon deformable task Fold Cloth on a real Piper arm (vs.20.0% for π0\pi_0).

Keywords

Cite

@article{arxiv.2605.07931,
  title  = {One Token Per Frame: Reconsidering Visual Bandwidth in World Models for VLA Policy},
  author = {Zuojin Tang and Shengchao Yuan and Xiaoxin Bai and Zhiyuan Jing and De Ma and Gang Pan and Bin Liu},
  journal= {arXiv preprint arXiv:2605.07931},
  year   = {2026}
}
R2 v1 2026-07-01T12:58:05.208Z