Reproduce “Progress Measures for Grokking on Real-world Tasks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Satvik Golechha
This research demonstrates that grokking, a phenomenon of delayed generalization, can occur in real-world datasets even when the L2 norm of model weights continuously increases, directly challenging a prevalent hypothesis. The study introduces and validates three novel progress measures—activation sparsity, absolute weight entropy, and approximate local circuit complexity—that consistently track the onset of generalization across various training conditions and model architectures.
Key insights
- Grokking can manifest independently of decreasing L2 weight norms; generalization was observed on MNIST even when the L2 norm continuously increased during training.
- The proposed measures—activation sparsity, absolute weight entropy, and approximate local circuit complexity—consistently correlate with the onset and progression of generalization during grokking.
- Dropout regularization prevents delayed generalization, suggesting a direct influence on the model's internal circuit complexity and robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.