MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training
Lucia Gordon, Serge Belongie, Christian Igel, Nico Lang, [0.5em] Harvard University, USA · 2026
Read MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training by Lucia Gordon, Serge Belongie, Christian Igel, Nico Lang, [0.5em] Harvard University, USA (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-2602-06285.txt.
Description
Satellite and airborne Earth data come in many forms, from photos to radar to elevation, yet models usually train on a few forms and fail far from home. Lucia Gordon and 3 co-authors introduce MMEarth-Bench, five environmental tasks with a dozen data forms and global coverage. Test splits are both random and split by geography.
Benchmarking many pretrained models, they found pretraining helps when labels are scarce, but geographic transfer stays poor. A plain model with enough labels stayed competitive. Their fix uses every data form available at test time as extra training tasks, adapting the model on the spot. That method improved all models on both splits, and grouping by geography balanced fitting against overfitting.
Catalog record
| CREATOR | Lucia Gordon, Serge Belongie, Christian Igel, Nico Lang, [0.5em] Harvard University, USA |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 114 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-06285