When and Where to Reset Matters for Long-Term Test-Time Adaptation

Researchers propose an Adaptive and Selective Reset (ASR) scheme to address model collapse in long-term test-time adaptation, where AI models gradually degrade and predict only a few classes. The solution dynamically determines when and where to reset models while preserving beneficial knowledge through importance-aware regularization.

• →Long-term test-time adaptation causes model collapse where AI systems predict only a few classes for all inputs due to accumulated errors.
• →Current periodic reset strategies are suboptimal as they occur independently of actual collapse risk and cause catastrophic knowledge loss.
• →The proposed ASR scheme dynamically determines when and where to reset based on actual risk assessment.
• →An importance-aware regularizer helps recover essential knowledge lost during resets.
• →Extensive experiments demonstrate effectiveness particularly under challenging domain shift conditions.