Figure 4

Where the model wants to click

The 256-bin x and y slots, read in one pass and drawn back onto the screenshot. The model was trained to put a narrow Gaussian bump of probability over the right columns, not all of it on one, so its answer is a peak with a little spread onto neighbouring columns. Hover the screenshot to look at the columns: each is a few pixels wide and has its own token.

P(x) →
P(y) ↓

zoom · the columns under the cursor