A frozen teacher model improves test-time adaptation robustness.

Computer models often face shifting data in the real world without correct answer labels. To adjust, many systems use a teacher-student setup where a guide model directs a learner model. People usually update the teacher slowly as the student learns. However, researchers found that this method still gathers mistakes over time, especially during long tasks. To fix this, they tested an intransigent teacherintransigent teacherA teacher model whose weights are kept fixed and not updated during adaptation. that never updates its weights at all. This guide stays fixed while the student continues to adapt. Surprisingly, keeping the teacher frozen led to big performance gains across several datasets on longer scenarios. It also made the models less sensitive to hyperparameter tweaks and worked smoothly across different model designs, including semantic segmentation.

Why it matters

AI models often need to adapt to new data after deployment, like in self-driving cars or medical diagnosis. If errors build up over time, it can be dangerous or unreliable. This work shows that keeping the teacher model unchanged prevents these errors, making AI more stable in changing real-world environments.

Questions answered