Attention vs. Attribution

Attention asks where a component reads. Attribution asks how input features relate to the output. Those questions can point to different tokens.

For a sentiment prediction, a head may attend to sentence structure while integrated gradients assigns most output relevance to “excellent.”

Two explanatory targets

Attention: normalized internal mixing weights between positions. Attribution: an estimate of how inputs contribute to or affect a selected output. Agreement is informative; disagreement is also informative.

Compare Routing with Output Evidence

Attention and integrated-gradient attribution disagree on this positive review. Remove a disputed token to test which signal better predicts output change.

TokenAttentionAttribution
The
movie
was
surprisingly
excellent

Positive sentiment

0.93

Largest measured drop

run a test

Disagreement is useful. It tells you that internal routing and output sensitivity are answering different questions.

Do not pick the method with the prettier heatmap. Decide whether you need to inspect routing, identify sensitive inputs, or test a causal mechanism.

When methods disagree, perturb the disputed tokens and inspect the relevant internal pathway.

Checkpoint

What should you do when attention highlights “nurse” but attribution highlights “won”?