visv-Bro/to-grok-grokking-reproduction-code

To Grok Grokking — independent reproduction code

Self-contained scripts for an independent reproduction of ICML 2026 paper #17708, To Grok Grokking: Provable Grokking in Ridge Regression.

• theory_audit.py: bounded-Rademacher finite-dimensional audit of Theorems 4.1, 4.2, and 4.4–4.6, including condition-relaxation controls.
• ridge_gpu_sweep.py: paper-scale spectral GPU reproduction of the Figure 2 weight-decay and sample-size panels.
• relu_gpu_sweep.py: declared-Gaussian reconstruction of the random-feature and fully trained two-layer ReLU experiments in Figures 3–4.
• make_poster_figures.py: high-resolution summary figures generated directly from the three result CSV bundles.

Primary manuscript: https://arxiv.org/abs/2601.19791v4

The paper supplies no official experiment code or raw data. These scripts are independent implementations and preserve the paper's asymmetric loss convention: training loss is half-MSE; population loss is MSE.

Reproduced outcomes

• Theorem audit: all 32 zero-teacher runs and all 96 nonzero realizable-teacher runs exhibited the expected delayed-generalization sequence; all selected Theorems 4.1–4.6 inequality checks passed.
• Figure 2: 20 seeds at m=1000; log-log t2 versus weight-decay slope -1.000002, with all printed Equation (8) checks passing.
• Figures 3–4: all 18 fully trained ReLU runs grokked, with decay slope -0.99535; none of the six faithful random-feature runs reached test MSE 0.01 under the declared Gaussian-input protocol.

Executed GPU jobs:

• Figure 2: https://huggingface.co/jobs/visv-Bro/6a6f099f6b79c09949c1f39f
• Figures 3–4: https://huggingface.co/jobs/visv-Bro/6a6f09b06b79c09949c1f3a1
• Raw outputs: https://huggingface.co/buckets/visv-Bro/to-grok-grokking-results