Revisiting the Platonic Representation Hypothesis: An Aristotelian View | arXiv 

The similarity scores PRH reports are uninterpretable without a baseline, and their null (chance) baseline isn't zero: it grows with model width and depth, the very properties that scale alongside the capability PRH credits for the rising alignment. The paper's fix is to break the pairing between the two representations by shuffling rows, recompute the similarity, and repeat many times to build an empirical null. The observed score is then judged against that null (now you have p-value and threshold). The finding is that the cross-modal convergence in global metrics (CKA, Procrustes) largely vanishes, while local neighborhood agreement (mKNN) survives nearly intact.

--