At-Grok Is Not Converged:A Measurement-Validity Audit for Grokking Representation Metrics

2026-07-07 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

"At-Grok Is Not Converged: A Measurement-Validity Audit for Grokking Representation Metrics" reveals that a network's embedding compression on modular arithmetic tasks significantly lags its generalization, by at least 10,000 steps. The study found that reading effective rank at the grokking transition overstates the converged value by 3-5x for an MLP and 1.3-1.5x for a transformer, also obscuring compression in MLP cells. A key ablation showed LayerNorm's impact, shifting the fraction of compression completed by the grok step from 0.87 to 0.25 in a transformer, with scale invariance ruled out as the cause. The authors released an audit package designed to separate onset from compression, flag censoring, exclude non-generalizing boundary cells, and verify reference floor plateau, which identified a false-confidence bug in their own system. Additionally, an MLP-specific depth law linking norm budget to converged floor failed a generality test on transformers and reversed sign under free weight decay. Code and the toolkit are publicly available.

Key takeaway

For Machine Learning Engineers evaluating model generalization, you must account for the significant lag between accuracy gains and true embedding compression. Relying solely on effective rank at the grokking transition will likely overstate actual convergence by 1.3-5x, leading to misinterpretations of model learning. Integrate the provided audit toolkit into your analysis workflow to accurately assess representation dynamics, ensuring your models have genuinely converged and compressed their representations, especially when using LayerNorm.

Key insights

Network embedding compression significantly lags generalization, overstating effective rank at grokking transition.

Principles

• Effective rank metrics can misrepresent true compression.
• LayerNorm significantly influences compression lag dynamics.
• Grokking metrics require careful validation for convergence.

Method

An audit method separates onset from compression, flags censoring, excludes non-generalizing boundary cells, and verifies reference floor plateau for grokking metrics.

In practice

• Use the released audit toolkit for grokking analysis.
• Validate compression metrics beyond initial generalization.
• Consider LayerNorm's effect on representation dynamics.

Topics

• Grokking
• Representation Learning
• Neural Network Compression
• Effective Rank
• Layer Normalization
• Measurement Validity

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Deciphering Two Training Clocks in Grokking via Deep Linear Network Theory with Conditional ReLU Reduction — Grokking arises from two distinct training clocks: fast classifier fitting and slow representation simplification.
• SFT Overtraining Predicts Rank Inversion via Entropy Collapse Under RLVR — SFT overtraining causes entropy collapse, leading to GRPO rank inversion and failure, identifiable by pre-RL entropy.
• Quantifying Hyperparameter Transfer and the Importance of Embedding Layer Learning Rate — Maximal Update (\$μ\$P) improves LLM hyperparameter transfer by maximizing the embedding layer's learning rate, a bottleneck in standard parameterization.
• I-SAFE: Wasserstein Coherence Metrics for Structural Auditing of Scientific AI Models — Scientific AI model performance benchmarks are insufficient to confirm alignment with domain-relevant structural behavior.
• Representation Gap: Explaining the Unreasonable Effectiveness of Neural Networks from a Geometric Perspective — Neural network generalization is geometrically characterized by the Representation Gap, scaling with intrinsic data dimension and model equivariance.
• What Does the Weight Norm Control in Grokking? Logit-Scale Mediation under Cross-Entropy — Weight norm controls grokking delay primarily through logit scale and softmax saturation, especially under cross-entropy.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.