Self- Soup ervision: Cooking Model Soups without Labels
Anthony Fuller, James R. Green, Evan Shelhamer · 2026
Read Self- Soup ervision: Cooking Model Soups without Labels by Anthony Fuller, James R. Green, Evan Shelhamer (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-2602-02890.txt.
Description
Anthony Fuller and 2 co-authors extend a training trick to unlabeled data in a 2026 paper. The trick, called model soups, averages the weights of several fine-tuned models into one better model. All earlier soups needed labeled data and the same training objective. Their version works with self-supervised learning instead.
Self-souping can use new unlabeled sources, such as data from a shifted setting. Souping on corrupted test data, then fine-tuning on clean training data, improved robustness on corrupted image benchmarks. Ingredients could even use different self-supervised algorithms, a first for soups. Mixed soups of four different methods beat every single ingredient model on accuracy.
Catalog record
| CREATOR | Anthony Fuller, James R. Green, Evan Shelhamer |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-02890