DiVeR: Decision-Critical Verifier Learning for VLA Test-Time Scaling

DiVeR. Decision-Critical Verifier Learning for VLA Test-Time Scaling. By Seongheon Park and seven others, from the University of Wisconsin-Madison, Microsoft Research Asia, Tokyo, and three other institutions. Published on the fourth of October. No code repository is linked to it.

Vision-Language-Action policies have improved by scaling robot data and model capacity, but robot data is expensive, and that cost limits how much further the scaling can go. Verifier-guided test-time scaling is an efficient alternative. At inference time, the policy samples several candidate actions, and a verifier picks the one most likely to lead to task success. Existing verifiers of the classification kind learn from outcomes at the level of whole trajectories, and they treat every state the robot visited as equally informative. That is the flaw. At many states the plausible actions are much alike, and there is little to discriminate between. Only a sparse set of decision-critical states admits actions that are meaningfully different and can substantially change what happens next.

DiVeR estimates how decision-critical a state is from the dispersion of the action representations sampled there. When sampled actions spread widely, the choice among them matters. When they cluster, it hardly does. That signal is then used to reweight verifier learning toward the states where action selection is most consequential.

The appeal is in what it does not need. Criticality comes from the policy's own samples, so DiVeR requires no step-level annotations, and no additional interaction with the environment. The verifier is trained on the same kind of data as before, with its attention moved to the moments that count.

DiVeR is tested on LIBERO, on RoboCasa, and in real-world experiments with a Franka Research 3 robot. Across all three it consistently improves task success through more effective verifier-guided action selection. The verifier's inference overhead grows by a negligible amount.

A trajectory's outcome is shared by every state along it, but only a few of those states were real decisions. DiVeR uses the spread of the policy's own samples to find them, and teaches the verifier to look hardest there.