Revisiting the Platonic Representation Hypothesis: An Aristotelian View

The Platonic Representation Hypothesis claims that as models scale, their representations across architectures and modalities (vision, language, video) increasingly converge. We revisit this claim and show that widely used similarity metrics can be systematically inflated by model scale (width and depth). After correcting for these effects with a permutation-based calibration, a more nuanced picture emerges: global spectral “convergence” mostly disappears, while local neighborhood relationships remain robust.