A new paper on arXiv (2605.20337) takes up a pressing question: how interpretable are the features of leading vision models? The authors frame this as increasingly urgent because these models are moving from research benchmarks into high-stakes deployments.
The paper argues that existing methods cannot answer this question reliably. That gap matters, the authors suggest, because a model's capability does not necessarily tell us how interpretable it is.
The title itself states the central claim: capability is not interpretability. The work points to the need for better ways to evaluate what these models actually show us.