Gate Provenance Determines Reliability of Training-Free VLA Token Skipping

A new study posted on arXiv (2608.00391) investigates how reliable training-free token skipping is in vision-language-action (VLA) models, which is meant to enhance inference speed by skipping calculations for many visual tokens during each control step. The findings reveal a major problem: when the skipping gate relies on the model's quick forward pass, skipped tokens become invisible to the next gate, leading to accumulated errors that can cause task failures. The researchers experimented with two strategies—reuse and deletion—using the same gate signal on the same episodes. At a 0.9 skip ratio on the LIBERO-Object benchmark, both methods performed poorly, with success rates falling to 0.68 and 0.31, respectively, compared to a perfect baseline of 1.00. Notably, the action-level detectors used in the research don’t catch this failure. The key takeaway is that the effectiveness hinges on whether the gate is generated from a complete forward pass, rather than the approach itself. This paper is categorized as a cross announcement and is available on arXiv.