CARE: Certifying Acceleration for Vision-Language-Action Inference
CARE picks the fastest VLA accelerator it can certify without exceeding a user-set failure budget.
The paper defines failures by paired rollouts: same initial condition, reference policy succeeds, accelerated policy fails. CARE tests candidate accelerators on a calibration set, then deploys the fastest one with a finite-sample guarantee. On LIBERO with OpenVLA-OFT, it reports 9.0-10.8x certified speedups while preserving at least 85.8% of reference-solved episodes at 95% confidence. The authors also report lower rollout cost from sequential testing and show the method transferring to flow-step reduction and language-agent settings. HF Daily Papers' note
The paper defines failures by paired rollouts: same initial condition, reference policy succeeds, accelerated policy fails. CARE tests candidate accelerators on a calibration set, then deploys the fastest one with a finite-sample guarantee. On LIBERO with OpenVLA-OFT, it reports 9.0-10.8x certified speedups while preserving at least 85.8% of reference-solved episodes at 95% confidence. The authors also report lower rollout cost from sequential testing and show the method transferring to flow-step reduction and language-agent settings. HF Daily Papers' note
score 4