固定
How can we reliably measure whether a representation identifies latent factors? Can current identifiability metrics evaluate interpretability or unsupervised mechanism discovery?
Our UAI 2026 paper finds systematic false positives and false negatives in these evaluations.