Robust calibration for dependent, anisotropic, high-dimensional hidden-state embeddings

Develop more robust calibration procedures for the ASK-NN asymmetric nearest-neighbor test when applied to real hidden-state embeddings exhibiting dependence, anisotropy, and high dimensionality, potentially using permutation or block-resampling methods.

Background

The paper concludes that the asymptotic calibration of ASK-NN may be inaccurate for real hidden-state embeddings because token representations can be dependent, anisotropic, and high-dimensional. It identifies permutation or block-resampling calibration as possible alternatives, but does not establish a theoretically valid or empirically optimal solution for this setting.