CARE: Certifying Acceleration for Vision-Language-Action Inference

CARE. Certifying Acceleration for Vision-Language-Action Inference. By Rui Liu and three others, from the University of Maryland, College Park, the University of Oxford, and one other institution. Published on the sixth of October.

Vision-language-action models have advanced quickly, but running one at every control step is expensive. Earlier work speeds up inference with techniques like action chunking and visual-token pruning, and usually judges the result by latency and average task success. That average is where the risk hides. Acceleration can discard information and break tasks the original policy would have solved, and a healthy mean can conceal those losses. Measuring them is hard for a specific reason. Small deviations in action compound over a closed-loop trajectory, so a failure only becomes visible across a full episode, not at any single step.

The paper first pins down what counts as a failure caused by acceleration. Two rollouts start from identical initial conditions, one with the reference policy and one with the accelerated version. An acceleration-induced failure is the case where the reference succeeds and the accelerated policy fails. With that definition in hand, CARE treats speed-up as a selection problem with a guarantee attached. It picks among candidate accelerators, but only from those it can certify.

Certification runs on a calibration set of paired rollouts. From them, CARE provides finite-sample guarantees that the risk of acceleration-induced failure stays below a budget the user specifies. It then deploys the fastest candidate that clears that bar, and if none qualify, it falls back to the reference policy. Because it looks only at terminal outcomes and measured compute, it does not need to know how an accelerator works, so the same procedure applies unchanged across very different acceleration mechanisms. Two devices keep the cost of certifying down: sequential testing, and reference rollouts that are triggered only by failures.

The main tests use OpenVLA-OFT on four LIBERO suites. There CARE certifies speed-ups of nine to ten point eight times, while guaranteeing, at ninety-five percent confidence, that at least eighty-five point eight percent of the episodes the reference solves are preserved.

Tight budgets show the contrast most clearly. Selectors without guarantees exceed the budget in up to seventy-five percent of trials. CARE stays within budget, and its sequential form uses seventy-eight point nine percent fewer rollouts than exhaustive evaluation. The approach also carries over to flow-step reduction for pi zero point five, and to Qwen3.5-9B and Llama-3.1-8B agents in Crafter.

CARE changes the question asked of a faster policy. Instead of whether it succeeds about as often on average, the question becomes whether it keeps the specific episodes the original would have won, with a bound that holds at a stated confidence. Since the check needs only outcomes and compute, any accelerator can be put through it, and the slow reference remains the safe default whenever nothing passes.