acceptodds
Under review as a conference paper at ICLR 2027

Data Coverage Gates the Grokking Transition in Small Transformers

Abstract

Grokking (sudden generalization long after memorization) is well-studied in small transformers on modular arithmetic, but the role of training-data diversity in the transition is unclear. In this controlled setting we isolate unique-pair coverage (the fraction of the p² ordered pairs seen) from total sample count. Our central finding is that unique-pair coverage is a causal, hysteretic boundary for the formation, but not the maintenance, of the Fourier circuit that enables grokking. Raising coverage past the critical value c* mid-training triggers grokking (9/10 seeds vs. 0/10 control); cutting it below c* after the circuit forms leaves it intact (10/10); and far below c* the failed state is operatively absorbing over 500,000 steps. Where the boundary is: within a budget of 100,000 steps, c* ≈ 24% of unique pairs at the default hyperparameters (15.6%–34.9% across weight-decay and learning-rate settings), with c*(p) ∝ p^(−0.36) under the held-out criterion, an intermediate scaling of the required pair budget between the O(p log p) and O(p²) predictions. What crossing it changes: the representation stays diffuse below c* and collapses onto a sparse Fourier circuit only above it, a circuit-efficiency count whose predicted exponent falls inside the measured interval. What it is not: the threshold reproduces under the without-replacement epoch protocol of prior work, is not an artifact of repetition or frequency skew, and carries over within the family to a second task and a two-layer transformer of the same width. Because data-scaling experiments vary coverage and sample count together, dataset-size effects should be read with coverage held fixed.

Primary area:
interpretability and explainable AI
Keywords:
grokking, delayed generalization, data diversity, sample complexity, transformers, algorithmic tasks

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LccEOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper explores how the coverage of the training data effects grokking in two one and two layer transformers models that have a width of 128, with a fix training recipe. They find that the coverage of the data has an impact on the delayed generalisation, with a critical threshold of approximately 24% with 100,000 training steps. Above this threshold the model groks below it the model does not.

Critical strengths

It should be commended the authors use of statistical analysis and experimental design, for example the use of factorial design, to deduce the findings.

Critical weaknesses
  1. This paper misses a core citation [1], which shows that training datasets with higher coherence, increases the number samples needed to delay generalisation, whereas datasets with lower coherence results in faster generalisation, with fewer samples required. Where high coherence means that data is more similar or aligned and low coherence means the data is very different. Ultimately they showed that grokking can be mitigated through data selection. In the paper being reviewed the authors explore unique-pair coverage as the fraction of the $p^2$ ordered pairs seen, without regard for the coherence of the data, which may explain the soft boundaries observed. Furthermore, it is unclear to me what this paper is telling me beyond what is already known that data selection plays a role in grokking as shown concretely in [1].

  2. The evidence used to support the claims is only shown for modular addition and multiplication and on two types of transformer architectures, even though grokking has been exhibited on more complex tasks [3]. This hampers the bold claim of the paper stated by the title Data Coverage Gates the Grokking Transition in Small Transformers. For such a statement, I would expect an exploration of a range of transformer architectures and training hyperparameters, to ensure this affect is not dominated by a singular training recipe or optimisation procedure. For example similar to Figure 6 in [3], which explores the landscape of grokking under hyperparameters. It is important to note that this limitation is acknowledged our evidence is confined to small models on modular arithmetic: one- and two-layer transformers of a single width (dmodel = 128; ∼216K and ∼414K parameters) under one training recipe, so the behavior of c∗ at larger scale or on other task families is untested and we claim no transfer; but no rational behind the decision is provided. Given the computational cheap costs of exploring this problem, this limitation is too limiting. It is unclear if such analysis is generalisable or extends to other learning tasks, architectures, optimisation setups. I feel the community will not gain any meaningful insights from this work, nor would there be much interest due to the limited scope of exploration.

[1] Grokking Beyond the Euclidean Norm of Model Parameters Tikeng Notsawo Pascal Junior, Guillaume Dumas, Guillaume Rabusseau Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:28552-28618, 2025.

[2] Liu, Z., Michaud, E.J. and Tegmark, M., 2022. Omnigrok: Grokking beyond algorithmic data. arXiv preprint arXiv:2210.01117.

[3] Liu, Z., Kitouni, O., Nolte, N.S., Michaud, E., Tegmark, M. and Williams, M., 2022. Towards understanding grokking: An effective theory of representation learning. Advances in Neural Information Processing Systems, 35, pp.34651-34663.

Questions for authors

See weaknesses but more concretely:

  1. Could the authors situate there work in regards to [1], and explain the how it goes beyond there findings?

  2. Could the authors provide further analysis across a range of transformer architectures and training hyperparameters, to ensure this affect is not dominated by a singular training recipe or optimisation procedure?

  3. Could the authors extend the analysis to more complex learning tasks as found in [2]?

[1] Grokking Beyond the Euclidean Norm of Model Parameters Tikeng Notsawo Pascal Junior, Guillaume Dumas, Guillaume Rabusseau Proceedings of the 42nd International Conference on Machine Learning, PMLR 267:28552-28618, 2025.

[2] Liu, Z., Michaud, E.J. and Tegmark, M., 2022. Omnigrok: Grokking beyond algorithmic data. arXiv preprint arXiv:2210.01117.

Reviewer_xHfrOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies grokking in small Transformers trained on modular arithmetic, treating the fraction of distinct input pairs seen during training (unique-pair coverage) as the control variable. At p = 67 under a 100,000-step budget, a logistic fit places the 50%-success coverage threshold near 24%. The core experiments intervene on nested training supports mid-run: increasing coverage from 15% to 30% at step 25,000 triggers held-out generalization in 9/10 cosine-schedule runs (versus 0/10 in fixed-coverage controls), whereas reducing coverage from 40% to 20% after grokking preserves high final accuracy. Constant-learning-rate controls help separate learning-rate scheduling effects from support effects. The paper also characterizes the transition through Fourier-entropy tracking, readout-frequency ablations, sampling-protocol controls, and scans across moduli p, finding an empirical scaling of c*(p) ~ p^(-0.36) over the tested range.

Critical strengths
  1. The support-intervention protocol is carefully controlled. Section 4.2 and Appendix A.3 use nested subsets and evaluate strictly on pairs unseen in any training phase, addressing potential leakage in the support-removal experiment. The contrast between the 9/10 success rate under coverage injection and 0/10 in the fixed-coverage control, together with the constant-learning-rate ablations, provides clear evidence of how support changes affect optimization dynamics.

  2. The paper makes a thorough effort to test alternative explanations and document caveats in the appendix. These include a without-replacement epoch control, coverage-matched frequency-skew comparisons, held-out metric reanalyses, seed-level uncertainty estimates, an extrapolation check at p = 257, and a candid explanation of why the uniform-repetition experiment is distributionally redundant under i.i.d. sampling. Appendix A.10 also explicitly reports that a constant-kappa circuit-count model fails to capture the scaling trend, rather than obscuring that negative finding. Together, these controls make the empirical results credible.

Critical weaknesses
  1. The novelty of the findings beyond closely related prior work is not sufficiently established. Nanda et al. [2] already scan the fraction of distinct modular-addition pairs and observe non-generalization, delayed grokking, and immediate-generalization regimes. Varma et al. [1] explicitly study a critical dataset size and the effect of reducing the training set after generalization in their ungrokking experiments (Section 5.2). Their Appendix A.2 additionally demonstrates training-history-dependent data requirements: similar intermediate generalization occurs at roughly 2,000 examples in a semi-grokking run versus 800–1,000 examples when starting from a fully generalizing model, providing a direct qualitative precedent for different data requirements before and after learning. Ootani [3] also explicitly studies coverage-gated grokking and multi-seed success rates. The manuscript does not clearly establish what new scientific conclusion its combined controls and denser threshold measurements add beyond these prior results. At fixed p, coverage and unique sample count satisfy c = N_unique/p^2, so renaming the axis does not introduce a new control variable; likewise, uniform duplication leaves the i.i.d. sampling distribution unchanged. Consequently, the empirical contribution is largely incremental.

  2. The boundary and mechanistic claims are not sufficiently supported by the experiments. The 40%-to-20% removal experiment ends in a regime where some runs already grok from scratch and extended training rescues 4 of the 9 initially unsuccessful runs. Removal into the 15% regime, where all ten extended from-scratch runs fail, is not reported. Under a constant learning rate, accuracy in the removal arm drops to a mean post-swap minimum of 39.6% before recovering above 90% in 9 of 10 runs (Appendix A.3, Table 6). Thus, the evidence shows finite-time history dependence at the 20% support level, but does not establish distinct formation and maintenance boundaries, especially since transient collapse followed by recovery differs from uninterrupted maintenance.

The circuit-counting account also has limited independent predictive power. In Appendix A.10, kappa is defined from the observed threshold. Under the scaling parameterization c*(p) ~ p^(-alpha), holding kappa fixed while using the measured entropies yields alpha near -0.035 rather than the empirical 0.359. The frequency-based held-out prediction has an MAE of 4.1 percentage points across six moduli, compared with 0.74 for the empirical power-law fit. Moreover, the top-frequency ablation removes more update norm than its random control, and all ablated conditions still eventually grok (Table 16). Overall, the proposed threshold mechanism remains substantially less established than the descriptive findings.

Questions for authors
  1. What specific new conclusion is established beyond Varma et al. [1] (Section 5.2 and Appendix A.2) and the coverage studies in [2, 3]? Please distinguish a new finding from a different parameter setting or a more systematic collection of controls. In particular, what does the formation–maintenance claim add to earlier observations of training-history-dependent data requirements?

  2. Does a formed circuit survive a 40%-to-15% cut under a constant learning rate and extended post-cut training, when compared against a matched from-scratch control? Please define “maintenance” explicitly and reconcile it with the post-swap accuracy drop (mean minimum 39.6%) and 9/10 final retention in Table 6.

  3. Can the circuit-counting model predict held-out thresholds without fitting or defining kappa from the threshold itself, and does the top-frequency ablation remain distinctive under an update-norm-matched control?

Reviewer_ATzvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper argues that unique-pair coverage, not total sample count, decides whether a small transformer groks modular arithmetic. It estimates a critical coverage and shows mid-training interventions that move runs across it, and fits a scaling law in these. The experiments and the statistics are nice but my concern is that the central message is not new, and the paper presents it as if it were. the diversity of distinct training examples, rather than their raw count, governs generalization is close to a textbook point. Data augmentation exists precisely to increase coverage of the input space. Within grokking itself Power et al. (2022) and nearly every follow-up already parameterize the phenomenon by training fraction, which is unique-pair coverage. Dataset size and coverage are the same quantity, so the confound the paper sets out to resolve largely does not exist.

Critical strengths
  • The experiments are careful and well designed.
  • The intervention design is the most interesting part. Adding coverage rescues runs stuck in the memorizing state, and removing coverage after grokking does not destroy the circuit.
Critical weaknesses
  • The main message is not novel. It is well established that the diversity of distinct examples, not raw count, drives generalization.
    • In the bias–variance tradeoff, repeated samples add no information.
    • In sample-wise double descent, the effective number of distinct examples relative to capacity sets the regime.
    • Data augmentation exists precisely to increase coverage indirectly.
    • Grokking work since Power et al. (2022) already parameterizes the phenomenon by training fraction, which is coverage.
Questions for authors
  • What do you see as the contribution beyond the known training-fraction dependence of grokking? Would you reframe the paper around the maintenance, and position it to similar work?
  • How that maintenance relate to the recent work on ungrokking?
  • Does the scaling hold anywhere other than the modular addition task?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.