The awkward question in interpretability research is how you know an explanation is right.
The standard move is comparison: run your method alongside the established ones, show the saliency maps agree, and call the agreement validation. This is circular and everyone knows it. If the existing methods share a systematic error — and there's decent evidence they do — then agreeing with them is evidence of nothing. You've measured consensus, not correctness.
The work here started from that complaint. If you want to say a method identifies the features a network actually relies on, you need a case where you independently know what those features are. That means constructing the ground truth rather than hoping to recover it.
Multi-way, not one-at-a-time
The second problem is that most attribution treats features one at a time — how much did pixel i contribute — when the interesting behavior is interaction. A network can be insensitive to two features individually and highly sensitive to their conjunction. Rank them separately and both look unimportant. The explanation is not merely incomplete, it is actively misleading about how the model works.
Assessing multi-way explainability means asking whether a method recovers those interaction structures, which is a materially harder test and one that several popular methods do not pass cleanly.
What I'd say now
Interpretability is not a solved problem and I'd be suspicious of anyone selling it as one. What I took from this is narrower and I still hold it: a method's agreement with other methods tells you almost nothing, and the field improves roughly in proportion to how willing it is to build evaluations that can actually return a negative result.
This is the same instinct as the ELIZA argument I keep making — surface plausibility isn't evidence about mechanism. An explanation that looks reasonable to a human is exactly the thing you should distrust most, because looking reasonable to humans is what these methods are optimized to do.