SabaPivot's picture
|
download
raw
2.42 kB

Reproduction: What Makes a Strong Model?

Independent reproduction of ICML 2026 paper #23301 / arXiv:2606.01292 / OpenReview ykWN4LG9vE.

The authors did not link runnable code in the paper or arXiv source as of 2026-07-21. This bundle therefore implements the diagonal spectral mechanisms from the paper and a scaled real-architecture proxy independently.

Scope

  • scripts/core_repro.py: exact diagonal population dynamics for Theorems 1, 3, and 5, plus a Monte Carlo teacher-SGD / population-student experiment for weak-to-strong early stopping.
  • scripts/architecture_spectra.py: scaled Figure 2(c) proxy using 1,000 images from deedax/UTK-Face-Revised and pretrained ResNet18, ResNet50, and ViT-B/16 features.
  • scripts/spectrum_fit_diagnostics.py: log-log $R^2$ diagnostics for the persisted architecture spectra.
  • paper_source/: official arXiv v1 source used to recover formulas and experiment details.
  • outputs/: local smoke-test and downloaded Hugging Face Job results.
  • .trackio/logbook/: canonical challenge logbook.

The full UTKFace training panels and ImageNet-based Figure 3(c,d) were not rerun: the paper does not release code/checkpoints, and those panels require several full pretrained feature sweeps and training runs. The architecture experiment here is explicitly a scaled proxy, not a full replication.

Local rerun

uv venv .venv
UV_CACHE_DIR="$PWD/.uv-cache" uv pip install --python .venv/bin/python numpy
.venv/bin/python scripts/core_repro.py --output outputs/core --reps 64 --seed 23301

The real-architecture proxy is a PEP 723 script and can be run with:

uv run scripts/architecture_spectra.py --output outputs/architecture --samples 1000 --batch-size 32

Hugging Face runs and artifacts

All repository-specific generated files stay under this reproduction directory.

Xet Storage Details

Size:
2.42 kB
·
Xet hash:
1d9f746ed308b1c7d2a62c6e0b0fce50ab5d05e425404d940878b31b29cc0e32

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.