Multi-Way Representation Alignment
Akshit Achara King’s College London [email protected] Tatiana Gaintseva Queen Mary University of London [email protected], Mateo Mahaut Universitat Pompeu Fabra [email protected] Pritish Chakraborty Indian Institute of Technology, Bombay [email protected], Viktor Stenby Johansso · 2026
Read Multi-Way Representation Alignment by Akshit Achara King’s College London [email protected] Tatiana Gaintseva Queen Mary University of London [email protected], Mateo Mahaut Universitat Pompeu Fabra [email protected] Pritish Chakraborty Indian Institute of Technology, Bombay [email protected], Viktor Stenby Johansso (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-06205.txt.
Description
Neural networks trained apart on similar data seem to build similar inner representations, a claim known as the Platonic Representation Hypothesis. Akshit Achara and 7 co-authors study how to align three or more such models at once. Older alignment tools work in pairs, which grows costly and gives no single shared reference.
The team adapts a classic multi-shape alignment method to build one shared orthogonal space that keeps each model's inner geometry. For search tasks, though, strict shape matching is not ideal. Their corrected method adds a fix for direction mismatch after alignment. Tests showed better any-to-any retrieval across models, while keeping a usable common space for jobs like model stitching.
Catalog record
| CREATOR | Akshit Achara King’s College London [email protected] Tatiana Gaintseva Queen Mary University of London [email protected], Mateo Mahaut Universitat Pompeu Fabra [email protected] Pritish Chakraborty Indian Institute of Technology, Bombay [email protected], Viktor Stenby Johansso |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06205