---
title: Reproduction — Nonlinear Autoencoders Reveal Hidden Structure
emoji: 🧬
colorFrom: indigo
colorTo: pink
sdk: static
app_file: index.html
pinned: false
tags:
  - trackio
  - trackio-logbook
  - open-experiment
  - icml2026-repro
  - paper-wm3ABfhE7P
  - arxiv:2602.10680
---

# A Solvable High-Dimensional Model Where Nonlinear Autoencoders Learn Structure Invisible to PCA While Test Loss Misaligns With Generalization

Five claim-matched executions using the authors' pinned autoencoder and AMP code, exact population-loss quadrature, destructive controls, and two warning-strict byte-identical replays.

Run `python reproduce.py --output-dir replay` and `python validate_evidence.py`.