Move the map

A language model reads a short description of where three things are. We never change the text. Instead we edit two numbers inside the model, in the sentence that places one object, and ask it six questions. Click a spot on the map to move that object there.

Show where a probe reads each object from the activations (rough: it reads positions well on average, R² 0.74, but the answers don't rely on that copy)

What the model reads (unchanged)

How far to move it (inside the model)

not at all¼halfway¾all the way

The model's answers

Question	Correct here	Model, after the edit