Definition
Plain language
A colored overlay showing where a model was looking hardest in an image.
As stated in the literature
A visualization of attention weights or gradient-based saliency projected onto the input; used here to show that captioning passes distribute attention globally while question-directed passes concentrate on the queried region.
Also called: heatmap, heat maps
Why it matters: It shows where a model is actually paying attention, which reveals that a general description of an image can completely skip the one small detail a question depends on.
For example, asking a model "what color is the sign?" lights up a bright patch right on the sign, while asking it to describe the whole photo spreads the color across the frame.
Heard on the show
“They make a heat map — skills down the rows, tasks across the columns, green where a skill helped, red where it hurt.”Episode 151 — Why More Experience Made This AI Agent Worse, And How to Fix It