Gemma 3 can build a detailed spatial map of a text description and then completely ignore it when answering questions. That is not a bug. It is a glimpse into how large language models actually think, and it should change how we evaluate them.
An independent, preregistered study tested Gemma 3 (12B and 27B parameters) on simple text layouts like "A is 3 north of B, C is 2 east of A." Probes could read every object's position from the model's internal representations with high accuracy (R² 0.74), even when the positions required combining two sentences. That map existed. Yet when researchers erased it, the model's accuracy on spatial questions did not drop. The map was irrelevant to the answers. Instead, the model used a separate 2-D code per sentence that acted like a logic switch, not a coordinate system. When the researchers read answers directly off the unused spatial map, that external readout actually outperformed the model itself on chain questions: 80 percent versus 73 percent. The model had the right geometry stored but did not use it.
This matters because it reveals a fundamental mismatch between what we assume models do and what they actually do. We treat them like reasoning engines that build and consult internal models of the world. In this case, Gemma 3 built the model and then bypassed it, relying on a simpler pattern-matching shortcut. The same tendency shows up in other domains. For instance, when validating LLM outputs in DuckDB and dbt, the real challenge is not whether the model knows the answer but whether it uses the knowledge it has without hallucinating shortcuts. And as Google's Gemini 4 Hydrogen arrives, optimized for speed and agentic workflows, the temptation to take the fastest path rather than the most accurate one only grows.
The practical takeaway is direct: do not trust a model's internal representations as evidence of its reasoning. A model that encodes spatial relationships perfectly may still answer spatial questions incorrectly. This is not a Gemma 3 problem. It is a property of how transformer architectures learn to compress information. The map exists because it is statistically useful during training. The model ignores it during inference because the shortcut is cheaper. For anyone building on these systems, whether in coding, search, or data validation, the lesson is to test outcomes, not internals. The model's hidden map is a monument to what it could do, not a record of what it will do.
The study is preregistered, the code is open, and the demo is live. We should watch what happens when researchers probe other models for the same behavior, especially as chain-of-thought prompting becomes standard. If a model can hold a perfect internal map and still answer from a shortcut, what else is it ignoring?