Representation Alignment

Representation alignment asks when different models, modalities, or groups of users encode the same useful structure, and whether those representations can support transfer or cooperation. Researchers compare spaces with methods such as CKA, RSA, nearest-neighbor overlap, and Procrustes alignment, or train mappings with contrastive losses, distillation, and lightweight adapters. Applications range from vision-language models and brain-signal decoding to federated learning, diffusion generation, and transferring capabilities between language models. A central open tension is that similar geometry or high alignment scores do not necessarily mean models behave alike: agreement may hold for broad relationships but fail for fine details or useful information transfer.