Halt Fast!: an early-stopping method for Randomized Smoothing certified robustness

Halt Fast! proposes an anytime-valid meta-learning framework for certified robustness, letting Randomized Smoothing allocate computational resources adaptively per input. The framework uses a lightweight meta-learner to predict image-specific priors for sequential E-processes, reducing sample complexity by 20 times while maintaining strict statistical guarantees. It can also allocate computation adaptively according to application risk thresholds, which the authors say provides a path toward real-time safety-critical certified deployment.