TM-006Coordinates are recoverable from TerraMind's frozen features, latitude more accurately than longitude

Kamil Dybek, Jakub Harchut, Oskar Kuliński

SourceWhere Are We Now? Investigating Spatial Information in TerraMind's Latent Space

A picture's coordinates can be partly decoded from TerraMind's frozen internal representation by training probe models on top of its features. Across all 12 combinations of four model variants and three probe types, latitude was decoded more accurately than longitude.

Evidence
correlational
Key metric
latitude R2 exceeds longitude R2 in all 12 probe configurations, 4 model variants by 3 probe types; the source counts latitude and longitude predictions separately as 24 experiments
Caveat
The conclusion is based on three probe families; the authors state that a broader and more principled selection of probing models is needed to improve interpretation of the latent-space structure.
Model
TerraMind v1 tiny, v1 small, v1 base, v1 large
Concepts
Linear representation
Datasets
TerraMesh [eval]
Methods
Probing classifiers / MLP probing classifiers / Q16 classifier [primary], Ridge regression / Bootstrap ridge regression [primary], XGBoost [primary], Multilayer perceptron [primary], Control task with shuffled labels [validation]
Related findings
TM-007
Extraction
automatic-extraction