NeurIPS paper dissects the Platonic Representation Hypothesis and proposes pair-free NGMA alignment

loremstasum · reddit · 2026-11-23

A NeurIPS 2026 paper questions how the Platonic Representation Hypothesis is tested: nearly all evidence uses cosine similarity, a kernel only contrastive/JEPA models are trained to share, while in VAEs it reflects prior and loss choices. The authors propose NGMA (Neural Gromov–Monge Alignment), requiring no pairs or labels. Key claims: distances pulled back from data space are modeling-choice invariant; alignment is identifiable iff data break all manifold symmetries, and is the unique Gromov–Wasserstein optimum. Experiments: 87% test-image retrieval between two FashionMNIST VAEs with different priors (Rosetta: 6.5%, rigid map on true pairs: 7.1%); up to 38% of 10,000 held-out scVI lung cells correctly paired. Limitations: single modality, no large pretrained encoders, guarantees only for global optima (7 of 35 CelebA runs aligned faces to mirrors).

Original post →

More from Research

Research channel →