The Attention-as-Explanation Debate

One 2019 paper argued that attention is not an explanation. A rebuttal published that year argued that attention can still provide explanatory evidence. And the debate continues to this day!

The strongest criticism is that very different attention patterns can sometimes produce the same output.

Routing is not the final reason

Attention weights expose one internal operation. They can be weakly correlated with gradient attribution, may change without changing the prediction, and cannot always be intervened on while every other variable stays fixed.

Same Prediction, Different Attention

Switch between two attention distributions for the pronoun “she.” Both produce the same toy coreference prediction even though the internal routing looks different.

The
0.02
doctor
0.14
told
0.04
the
0.03
nurse
0.48
she
0.20
won
0.09

Coreference prediction

nurse

Prediction probability

0.82

Attention changed

baseline

Attention describes information routing inside one component. It does not automatically identify the input evidence that caused the final prediction.

Attention can still support an explanation when the claim is narrow: “this head routes information from the pronoun to the noun.” It is weaker evidence for “the model predicted this because of that noun.”

Match the visualization to the claim. Architecture inspection and output attribution are different jobs.

Checkpoint

When is attention evidence most defensible?