Gradient-based attribution methods tell you which model parameters matter for a particular input, but they require backpropagation for every sample, which quickly becomes expensive. Static importance scores, by contrast, are cheap to compute but give the same answer regardless of the input, discarding the conditional information that makes attribution useful.
The paper introduces Kernelized Parameter Sensitivity Atlases (KPSA) to address this trade-off. The method is derived as an amortized form of conditional interpretability, aiming to provide input-specific parameter sensitivity information without the per-sample cost of gradient backpropagation.
That is the core claim in the abstract; the full derivation and evaluation are not included in the available text.