XGBoost

anchor

Gradient-boosted decision trees. As a probe it sits between a linear probe and an MLP: it can capture non-linear structure without being a full neural network, so it is used to separate "not linearly readable" from "not readable by anything simple."

Findings

• IC-992LLaMA-2-7B's last hidden layer encodes the prompt format with high identifiability, and the separability of format embeddings in the top two principal components correlates with performance spread [primary]
• TM-006Coordinates are recoverable from TerraMind's frozen features, latitude more accurately than longitude [primary]
• TM-007Coordinates are linearly decodable only in the larger TerraMind variants [primary]