TM-007Coordinates are linearly decodable only in the larger TerraMind variants

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 read back out of TerraMind's internal representation by training small probe models on its frozen features. How easily that works with a simple linear probe depends on model size: in the small variants the information is present but tangled, so a non-linear probe is needed to reach it, while the bigger models lay it out more plainly. Ridge regression's average latitude error drops from about 1092 km on the Tiny model to 550 km on the Large one.

Evidence
correlational
Key metric
mean latitude error 1092 km (Tiny) to 550 km (Large), Ridge regression
Model
TerraMind v1 tiny, v1 small, v1 base, v1 large
Concepts
Linear representation, Scale-dependent behaviour
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-006
Extraction
manual-extraction