Work

The Model Knows Better Than It Answers: Gemma 3 Composes a Spatial Map Its Answers Never Read

Year: 2026
Type: preprint
Abstract: When people read "A is 3 units north of B; C is 2 units east of A", they can answer questions the text never answers, as if looking at a picture. We ask whether a language model builds such a picture, and whether its answers use it. On text-only layouts, Gemma 3 12B builds a map: linear probes read each object's position from its activations (R² = 0.74), including positions that require combining sentences. But the answers do not use this map. Erasing it everywhere costs no more accuracy than erasing random directions. Instead, the answers depend on a small 2-D code over the sentence that states each position. Swapping it moves 72% of the affected answers (83% in Gemma 3 27B; random: 0%), and it acts as a switch, not a coordinate. The answers also use the chain of relations only partly, so answering from the model's own unused map beats the model on chain questions (0.80 vs 0.73). The main tests were preregistered.
Language: English
Open Access status: green
OpenAlex ID: W7225903884
Cites:
Cited by:
Related to: 10