AI BriefingThe Geometric Blind Spot of Supervised Models
Key point
It shows a structural geometric blind spot in supervised learning and argues that adversarial training makes it worse.
Details
A paper titled Supervised Learning Has a Necessary Geometric Blind Spot has been released, with arXiv number 2604.21395.
As its core claim, it presents that a model minimizing ERM (empirical risk minimization) is forced to retain non-zero Jacobian sensitivity in every direction that predicts the training labels.
The authors explain that this limitation does not disappear with model capacity or dataset size, and is instead determined by the data distribution.
They also argue that PGD adversarial training can actually make clean-input geometry worse rather than improving it, connecting this claim to both experiments and theory.
A code repository has also been released alongside the paper: PMH.
