Geo-foundation models (GFMs) are typically evaluated by how well they perform on downstream tasks. According to a new arXiv paper, this reveals little about which physical features the models are actually sensitive to or use in prediction. That leaves developers with limited ability to test what the models have learned.
The paper introduces PiCo, a physics-informed concept probing method. It is designed to probe GFMs for sensitivity to physical features, offering a way to see what the models are using rather than just their final accuracy. The abstract is brief, but the motivation is clear: without such probing, developers cannot easily tell whether a model is relying on the right physical signals.
The preprint is available at arXiv:2610.20973.