Research map / Label-efficient and robust learning / Test-time adaptation
MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training
Lucia Gordon, Serge Belongie, Christian Igel, Nico Lang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add multimodal reconstruction test-time training (TTT-MMR, random batching) after joint training (ablation) | 8 pretrained models plus ConvNeXtV2A baselines · MMEarth-Bench random and geographic test | average rank / R2 / mAP: avg rank 2.9 -> 2.1 (ConvNeXtV2A, random split) | Improved |
| use geographic batching instead of random batching in TTT-MMR (ablation) | TTT-MMR · MMEarth-Bench (most models) | average rank: 2.1 -> 1.1 (ConvNeXtV2A, random split) | Improved |
| use geographic batching instead of random batching (ablation) | TTT-MMR on Galileo · MMEarth-Bench random split | average rank: 1.7 -> 2.1 | Hurt |
| use geographic batching for rare (long-tail) samples (ablation) | TTT-MMR · MMEarth-Bench regression tasks | residuals by sample frequency | Improved |
| finetune with multimodal rather than S2-only inputs (ablation) | TerraMind, Copernicus-FM, Galileo · MMEarth-Bench geographic split of soil tasks | performance | Mixed |
| domain-specific satellite pretraining (DINOv3 Sat vs Web) (ablation) | DINOv3 · MMEarth-Bench | performance | No clear effect |
| train multimodal ConvNeXtV2A from scratch instead of using pretrained multimodal models (ablation) | ConvNeXtV2A-MM · MMEarth-Bench random test, 100% train | average rank: ranks 1st at 100% data, 4th at 5% data | No clear effect |
| evaluate on geographic (Africa) test split instead of random split | finetuned pretrained models · MMEarth-Bench soil N, soil OC, soil pH, species | performance | Hurt |
| reduce training data to 5% (ablation) | all models · MMEarth-Bench soil OC | R2 | Hurt |
Closest papers by meaning
- GeoCore-9B: Towards Geo-Aware Generative Foundation Models in Earth Observation (NeurIPS 2026 · similarity 0.87)
- GeoMMBench and GeoMMAgent: Toward Expert-Level Multimodal Intelligence in Geoscience and Remote Sensing (CVPR 2026 · similarity 0.87)
- AnySat: One Earth Observation Model for Many Resolutions, Scales, and Modalities (CVPR 2025 · similarity 0.86)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.86)
- GeoLLM: Extracting Geospatial Knowledge from Large Language Models (ICLR 2024 · similarity 0.86)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
