Platonic Representation Hypothesis

The Platonic Representation Hypothesis asks whether models trained independently on different data or modalities recover common structure in the world, visible in their internal representations. Researchers test this with representational similarity analysis, centered kernel alignment, neighborhood comparisons, and studies of cross-modal transfer, model stitching, and causal effects across vision, language, science, and world models. Evidence often finds shared local relationships or abstract semantic structure, but measured convergence depends strongly on the metric, scale, layer, and calibration; matching geometry does not reliably guarantee shared coordinates, transferable physical variables, or reasoning. A central open question is whether robust relational agreement reflects a useful common representation or a coarse artifact of training and measurement, and which objectives can produce stronger functional alignment.