World action models (WAMs) are designed to jointly predict future visual states and robot actions, but they face a practical bottleneck: processing dense future-frame tokens takes time. That latency makes them less suitable for real-time robotics applications where speed matters.

To tackle this, the authors propose CF-Prune, a training-free framework for pruning future tokens. Because it requires no additional training, it could be dropped into existing WAMs more easily than methods that need fine-tuning or architectural changes.

The arXiv abstract gives only a high-level view of the approach; it does not specify how tokens are selected for pruning or what trade-offs in prediction quality might arise. As a preprint, the work has not yet undergone peer review, so the reported latency improvements should be read with that caveat in mind.