Flow-Support-Weighted Proposals for Test-Time Verified VLA Policies

Key points

• A vision-language-action model trained to solve one task distribution concentrates its flow-matching decoder around the dominant demonstrated action, so repeated samples differ only marginally and the verifier receives many copies of one decision.1
• We call this proposal contraction and identify it as a key constraint on test-time scaling: repeated samples can remain highly redundant as the candidate budget grows.1
• Action Dynamic Sampling (ADS) alleviates the constraint on the proposal side, during training.1
• At each training step, ADS integrates $K$ additional trajectories from the current policy under fresh base noise, scores each by the flow support along its path with a Hutchinson trace estimator, and corrects the policy's local action distribution toward this weighted set through a detached finite-set KL term.1

Sources (1)

• [1]Flow-Support-Weighted Proposals for Test-Time Verified VLA Policies
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 05:06 AM

  Test-time sampling promises to turn one action prediction into a set of distinct executable candidates that a verifier can choose among.

  A vision-language-action model trained to solve one task distribution concentrates its flow-matching decoder around the dominant demonstrated action, so repeated samples differ only marginally and the verifier receives many copies of one decision.

Extractive summary: sentences quoted from the sources.