CARE Framework Provides Safety Guarantees for VLA Model Acceleration

CARE introduces a method for certified accelerator selection in vision-language-action models to prevent task failures caused by information loss during pruning or chunking. It uses paired rollouts on calibration sets to provide finite-sample guarantees against acceleration-induced errors.