@jedmaczan Great point! Gröger et al. (“An Aristotelian View”, ICML 2026) proposed a permutation-based null for calibrating such scores. We use their calibrated mutual k-NN too (small difference for local metrics) and add further references (Fig. 5): random init, VLMs, and unimodal pairs.
Do vision and language models, trained on entirely different data, converge toward a shared representation of the world? The Platonic Representation Hypothesis suggests they do. We take a closer look.
@AIQuanting Agreed, some convergence should come from shared content alone, so the raw number on its own says little. We put it in context (Fig. 5): randomly initialized models give the floor, and VLMs trained for cross-modal alignment give a reference for what “aligned” looks like.
@Grsdfd2 Great point, and exactly why we added a contextualization (Fig. 5 in the paper). We compare cross-modal alignment against three references: randomly initialized models as a lower bound, VLMs trained explicitly to align the two modalities, and unimodal pairs.
This work started as part of my bachelor’s thesis. Many thanks to my co-authors @SimonRoschmann, @QBouniot, and @zeynepakata for their amazing supervision throughout my thesis and beyond!