The Wittgensteinian Representation Hypothesis: Language Is the Attractor of Multimodal Convergence

arXiv·high signalXBlueskyLinkedInCopy link

Introduces directional convergence analysis using asymmetric cycle-kNN (invisible to standard symmetric measures) across dozens of independently trained unimodal models spanning point clouds, vision, and language. Finds a consistent directional asymmetry: non-language modalities converge toward language representations significantly more than the reverse, traced to feature density asymmetry — language occupies the most compact regions of representational space. Challenges the undirected Platonic Representation Hypothesis.