Model Merging by Output-Space Projection
2026
Read Model Merging by Output-Space Projection (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-2605-29101.txt.
Description
Bethan Evans and colleagues frame model merging as a convex quadratic programme over residual updates, replacing heuristics with a principled setup. Their method picks weights minimizing a squared-output calibration objective from calibration data, subsuming task arithmetic, model soups, TIES, and DARE as special cases. A closed-form diagnostic, the share of residual energy a chosen basis captures, predicts merge quality from calibration alone. It matches or beats methods in the single-layer case, and a sequential layer-wise version gains across language and vision benchmarks.
Catalog record
| YEAR | 2026 |
|---|---|
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50421 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.29101