Layer sensitivity analysis measures how quantizing or perturbing individual layers affects model behavior or accuracy. Run it before choosing a recovery strategy: diffuse failure points toward QAT, while failure concentrated in one layer points toward mixed precision. section 5.6
Diffuse versus concentrated failure
At INT4, ResNet-18 loses 12 percentage points across layers, while BERT-base falls to 50.9% accuracy because of a single FFN layer. These patterns motivate different interventions. section 5.6
