CASTER: Gradient-Free Test-Time Adaptation for Frozen Models
August 30, 2026
CASTER enables test-time adaptation for frozen or inference-only models without requiring gradients, optimizers, or backward passes. It uses affine statistics transport to estimate class-shared transformations from target-batch moments, making it viable for memory-constrained or third-party model deployments.
HOW THIS AFFECTS YOU
●
builderYou can implement domain adaptation on edge devices or with frozen APIs where backpropagation is impossible.
●
researcherYou can adapt models to new domains without needing access to model weights or gradient computation.