@misc{indiciaed8fd068eb27b,
  title = {A solvable high-dimensional model where nonlinear autoencoders learn structure invisible to PCA while test loss misaligns with generalization},
  author = {Vicente Conde Mendes and Lorenzo Bardone and Cédric Koller and Jorge Medina Moreira and Vittorio Erba and Emanuele Troiani and Lenka Zdeborová},
  year = {2026},
  url = {https://arxiv.org/abs/2602.10680},
  note = {Source identifier: 2602.10680}
}