Revisiting the Platonic Representation Hypothesis: An Aristotelian View

The Platonic Representation Hypothesis — the fashionable claim that as models scale they converge on one shared internal "statistical model of reality" — may be mostly a measurement artifact.

In Revisiting the Platonic Representation Hypothesis: An Aristotelian View, Fabian Gröger, Shuo Wen, and Maria Brbić (EPFL) show that the similarity metrics behind that claim are confounded by network scale. Wider models with higher embedding dimensions show a positive similarity baseline even when their representations are completely independent, and deeper models introduce a selection bias — reporting the maximum similarity across many layer pairs creates a "look-elsewhere" effect that inflates the numbers.

Their fix is a permutation-based null-calibration framework that turns any representational-similarity metric into a calibrated score with statistical guarantees. Once calibrated, the global-convergence signal behind most published evidence for the hypothesis "largely disappears." In its place the authors propose an "Aristotelian" view: what models actually share is local topological structure, not a single global representation.

Grigory Sapunov summarized the result on X: the Platonic Representation Hypothesis is "mostly a statistical illusion" — once calibrated, global convergence vanishes.

Paper: Revisiting the Platonic Representation Hypothesis: An Aristotelian View (arXiv:2602.14486).

@che_shr_cat on X →