The paper trains linear ridge regression probes on the internal activations of Llama-2 (7B, 13B, 70B) and Pythia (160M to 6.9B) to predict latitude/longitude of places or timestamps of events. Probes achieve high out-of-sample R² (e.g., 0.911 for world locations in Llama-2-70B), and representation quality increases smoothly through the first half of layers before plateauing. Nonlinear MLP probes yield minimal improvement over linear probes, confirming linearity. The representations are robust to prompting variations and unified across entity types (cities, landmarks, historical figures, etc.).
Evidence
correlational
Key metric
Llama-2-70B linear R² at 60% depth: world 0.911, usa 0.864, nyc 0.359, historical 0.835, entertainment 0.885, headlines 0.746; corresponding MLP R²: 0.926, 0.869, 0.312, 0.839, 0.884, 0.739
Caveat
High probe accuracy does not by itself imply the model uses these representations for its own predictions; the authors note this limitation explicitly and address it with neuron ablations in a separate experiment.