Flow-based Vision-Language-Action (VLA) policies are inherently stochastic: different noise seeds can produce action chunks of varying quality from the same observation. At test time, however, the policy typically commits to a single output, ignoring this variability. A new arXiv paper, VeriVLA, tackles this limitation by introducing adaptive test-time verification for flow-based VLAs.

The proposed approach leverages latent world models to verify action chunks before execution. Instead of blindly committing to one stochastic sample, VeriVLA adaptively checks the quality of candidate actions, aiming to select or refine the best one. This adds a verification step at inference time, which could improve reliability in real-world tasks.

The paper is available as arXiv:2610.21985v1. As the abstract is truncated, the full details of the verification mechanism and experimental results are not covered here, but the core idea is clear: use learned world models to make stochastic policies safer and more robust at deployment.