Grokking From Abstraction to Intelligence

Abstract: Grokking in modular arithmetic has established itself as the quintessential fruit fly experiment, serving as a critical domain for investigating the mechanistic origins of model generalization. Despite its significance, existing research remains narrowly focused on specific local circuits or optimization tuning, largely overlooking the global structural evolution that fundamentally drives this phenomenon. We propose that grokking originates from a spontaneous simplification of internal model structures governed by the principle of parsimony. We integrate causal, spectral, and algorithmic complexity measures alongside Singular Learning Theory to reveal that the transition from memorization to generalization corresponds to the physical collapse of redundant manifolds and deep information compression, offering a novel perspective for understanding the mechanisms of model overfitting and generalization.

• The paper shows that grokking arises from an abrupt collapse in neural network complexity driven by Occam's razor principles.
• It employs Singular Learning Theory and spectral diagnostics to reveal how redundant high-dimensional representations compress into essential subnetworks.
• Empirical analysis with a 48-layer Transformer on modular arithmetic validates the phase transition using metrics like BDM complexity and the Gini coefficient.

Structural Simplification and the Mechanistic Origins of Grokking

Problem Formulation and Limitations in Existing Explanations

The phenomenon of grokking—delayed generalization after overfitting, particularly in modular arithmetic domains—has become a central testbed for probing the inductive biases and internal mechanisms of overparameterized neural systems. Previous research has largely remained descriptive, focusing on task-specific circuit analysis, sensitivity to optimization and regularization parameters, or interpretability via local architectural probes. Such approaches have failed to capture the global, structural origins of the delayed generalization phenomenon, leaving unresolved the precise drivers and thermodynamic nature of the transition from memorization to generalization. Existing frameworks either lack predictive power, do not generalize across architectures, or sidestep the fundamental complexity dynamics underlying the emergent behavior.

Unified Theoretical Framework: Simplicity and the Occam Gate

This work proposes that grokking is a canonical instance of spontaneous structural simplification governed by the principle of Occam's razor, where the implicit regularization in large models drives an architectural collapse consistent with both Kolmogorov Complexity (KC) and the Minimum Description Length (MDL) principle. Generalization is characterized as a phase transition where the effective complexity of the solution manifold precipitously drops, as formalized through Singular Learning Theory (SLT) via the Real Log Canonical Threshold (RLCT). The critical claim is that model overcapacity allows a redundant high-entropy solution during memorization, which is later abandoned in favor of a parsimonious, algorithmically simple subnetwork aligned with the task’s algebraic structure.

Empirical findings demonstrate that after achieving perfect training accuracy, the model undergoes a sharp reduction in effective parameter utilization. This is evidenced by the collapse of algorithmic complexity proxies such as Block Decomposition Method (BDM), the concentration of representation energy in sparse spectral modes (Gini coefficient and IPR metrics), and the emergence of group-theoretic, low-rank manifolds in the learned embedding space. The model's parameter space initially supports a high-dimensional, functionally redundant solution, but transitions rapidly into a minimally sufficient circuit consistent with the target operation (e.g., modular addition's canonical representation as a circulant or diagonal structure in the frequency domain).

Empirical Analysis: Layer Bypassability, Spectral Localization, and Causal Mediation

The analysis is anchored by training a 48-layer Transformer on modular arithmetic, with performance and internal structure tracked at multiple checkpoints. Causal Mediation Analysis (CMA) via activation patching reveals that only certain early and late attention heads are necessary for function post-grokking; most intermediate layers become bypassable without loss of accuracy, demonstrating structural degeneration. This layer-wise bypassability is confirmed via systematic ablation, with negligible impact when removing intermediate blocks, supporting the assertion that the operational network's depth shrinks to a shallow, functionally-critical subgraph.

Spectral diagnostics uncover a synchronized reorganization: as memorization recedes, representation space collapses from a high-entropy diffuse manifold (visualized via PCA and persistent homology) to a 1D ring isomorphic to the cyclic group $\mathbb{Z}_p$, with the embedding matrix’s spectral density localizing in sparse, dominant Fourier modes. The Gini coefficient and IPR sharply ascend at the emergence window, aligning with the collapse of BDM complexity, while ablation and randomization controls eliminate weight decay or initialization as sufficient explanations for this localization.

Finally, quantification of algorithmic complexity by BDM—validated against nuclear norm, stable rank, and entropy baselines—shows a dramatic drop coinciding with generalization. The relationship between per-layer complexity drop and causal functional importance is linear ($R^2 \approx 0.88$), confirming that simplification is not only present but functionally necessary. These trends persist across model depths, modulus sizes, and tasks, establishing the universality of the mechanism in this domain.

SLT-Inspired Surrogate Modeling and Analytical Mechanisms

To bridge practical and theoretical insight, the Singular Feature Machine (SFM) is introduced as a surrogate model recapitulating the fit-complexity tradeoff underlying grokking. The SFM replaces opaque neural architectures with an analytically tractable machine operating on the Fourier-transformed input space, explicitly optimizing a free-energy-like objective combining data fit and an SLT-based sparsity penalty (proxy for RLCT). Training evolves in two phases: a memorization regime with maximal support in the spectral plane, and a critical crossover (modeled using the Occam Gate operator) to generalization, where the active support shrinks to a diagonal or otherwise group-aligned subset, effecting an $O(p^2) \to O(p)$ complexity reduction.

Complexity measures such as RLCT and KC are directly computable for the SFM, showing collapse at the same point as the empirical grokking event. Calibration and scaling analyses confirm that the theoretically derived critical sample size ($n^*$) correlates tightly with the empirical grokking onset time across various modulus sizes. Consequently, the SFM operationalizes the link between geometric (SLT), algorithmic (KC), and thermodynamic (free energy minimization) perspectives, providing a predictive and explanatory mechanism connecting overfitting, complexity collapse, and generalization performance.

Implications, Limitations, and Future Directions

The presented framework robustly grounds grokking as structural compression and spontaneous simplification in overparameterized systems, with strong support from cross-metric convergence, architectural ablation, null-control experiments, and surrogate theory. These findings imply that the generalization capability of large AI systems may fundamentally reflect implicit search for algorithmic minimality rather than incremental accumulation of explicit heuristic modules or circuits. This advocates for theoretical developments tying neural computation to SLT and formal complexity measures, and motivates experimental designs where model complexity, algorithmic compressibility, and functional specialization are tracked jointly.

Three key limitations are acknowledged: (1) The SFM, while analytically powerful, is not a strict derivation from actual Transformer training dynamics, but a hypothesis-generating construct; (2) Circuit analysis for non-additive modular operations (multiplication/division) lacks a fully worked-out mechanistic correspondence; (3) Additional baselines (e.g., varied optimizers, schedules, scales) would strengthen empirical positioning. Future research directions include extending this analysis to real-world algorithmic domains, explicit connection to information bottleneck principles, and cross-task generalizability to further test the universality of the identified phase transition in network structure and function.

Conclusion

This work establishes that grokking in modular arithmetic is not an incremental or heuristic process but a manifestation of an abrupt structural phase transition, where generalization arises from the spontaneous collapse of algorithmic and geometric complexity internal to the model. This process is robust, functionally necessary, and mechanistically aligns with formal principles of simplicity from SLT and KC. These insights contribute to a more unified foundation for understanding overfitting, generalization, and intelligence emergence in large-scale AI, with implications for both theory-driven analysis of deep learning and the principled design of complexity-aware architectures.

Reference: "Grokking From Abstraction to Intelligence" (2603.29262)