At-Grok Is Not Converged: A Measurement-Validity Audit for Grokking Representation Metrics
The study audits "grokking" representation metrics, revealing that embedding compression significantly lags behind accuracy generalization by at least 10,000 steps on modular arithmetic tasks. Standard metrics like effective rank at the grokking transition overstate converged values by 3-5x in MLPs and 1.3-1.5x in Transformers, failing to capture true convergence timing. Layer Normalization drastically reduces the fraction of compression achieved by the grok step (from 0.87 to 0.25), identifying
Analysis
TL;DR
- The study audits "grokking" representation metrics, revealing that embedding compression significantly lags behind accuracy generalization by at least 10,000 steps on modular arithmetic tasks.
- Standard metrics like effective rank at the grokking transition overstate converged values by 3-5x in MLPs and 1.3-1.5x in Transformers, failing to capture true convergence timing.
- Layer Normalization drastically reduces the fraction of compression achieved by the grok step (from 0.87 to 0.25), identifying architectural components as key drivers of compression lag.
- An MLP-specific depth law linking norm budget to converged floor fails generality tests on Transformers and flips sign under free weight decay, indicating limited cross-architecture applicability.
- The authors release an open-source audit toolkit designed to separate onset from compression, flag censoring, and exclude non-generalizing boundary cells to improve measurement validity.
Why It Matters
This research challenges the prevailing assumption that representation compression and generalization occur simultaneously during the grokking phenomenon, urging practitioners to decouple these metrics for accurate model diagnosis. By exposing significant biases in current measurement tools, such as overstated effective ranks and architectural dependencies, it provides a critical framework for validating theoretical claims about neural network learning dynamics. This ensures that future studies on emergent behaviors and representation learning rely on robust, audited metrics rather than potentially misleading proxies.
Technical Details
- Metric Discrepancy Analysis: The audit demonstrates that reading effective rank at the grokking transition overestimates the converged value by 3-5x for Multi-Layer Perceptrons (MLPs) and 1.3-1.5x for Transformers trained to convergence.
- Compression Lag Quantification: Embedding compression continues for tens of thousands of steps after accuracy has plateaued, with a lag magnitude comparable to the time-to-grok itself (minimum 10,000 steps).
- Architectural Ablation: Adding LayerNorm to an otherwise identical Transformer reduces the fraction of compression completed by the grok step from 0.87 to 0.25, while pre-registered controls rule out scale invariance as the underlying mechanism.
- Validity Audit Framework: The proposed toolkit separates onset from compression, flags censoring issues, excludes boundary cells that never fully generalize, and verifies that reference floors have plateaued.
- Generality Testing: A secondary finding regarding an MLP-specific depth law linking norm budget to converged floor was shown to fail when applied to Transformers and exhibited sign flipping under free weight decay conditions.
Industry Insight
- Researchers should adopt rigorous measurement audits that distinguish between accuracy transitions and representation compression to avoid misinterpreting model convergence timelines.
- Architectural choices, particularly normalization layers, significantly impact the temporal dynamics of representation learning; these effects must be accounted for when comparing model efficiencies across different designs.
- Open-source toolkits for auditing metric validity, such as the one released here, should become standard practice in deep learning research to ensure reproducibility and prevent false confidence in emergent behavior claims.
Disclaimer: The above content is generated by AI and is for reference only.