A new preprint (arXiv:2610.22749) takes up the Platonic Representation Hypothesis, the claim that larger models should learn increasingly similar representations of the same world. The authors point out that 'similar' is not a single thing: it can mean native cosine geometry, i.e. nearest-neighbour structure in the representation space, or it can mean linear recoverability, i.e. whether one model's representations can be recovered from another's via a linear map.
The paper's title indicates that alignment changes how these two notions scale with model size. The abstract excerpt provided does not include the results, so the direction and magnitude of that change are not summarised here.
For readers, the paper is a useful reminder that convergence between models is not a single phenomenon; different measures of similarity may diverge as models grow.