The linear representation hypothesis is a popular idea in interpretability: pretrained vision models encode generative factors as directions in feature space that can be read out linearly. The standard way to test this is to train a linear probe on the features and see whether the factor can be decoded.

A new preprint on arXiv, "Is a Linear Probe Evidence of a Linear Representation?" (2610.22655), pushes back on this routine. The abstract indicates that the paper questions whether a successful linear probe actually confirms the hypothesis, or whether the probe's flexibility makes the test too easy to pass.

Because this is a single preprint, there is no independent source to compare against. The full argument is not available in the abstract, but the title and abstract make clear that the burden of proof for linear representations is under scrutiny.