acceptodds
Under review as a conference paper at ICLR 2027

Activation-Spectrum Regularization for Grokking-Style Generalization

Abstract

Overparameterized networks can memorize a small algorithmic training set long before they generalize. We test whether compressing the spectrum of hidden-state covariance favors representations that reuse directions across examples. Likelihood-gated activation-spectrum regularization (ASR) applies a trace-based Rényi-2 proxy, spectral entropy, or ridge Log-Det only as predictive confidence grows. The Log-Det gradient weights an eigenvalue by the inverse of its ridge-shifted magnitude, and Sylvester's identity permits an exact computation through the smaller of the batch Gram and feature covariance matrices. ASR lowered 15 of 18 task–objective means and won 66 of 87 matched comparisons across the fixed six-task suite. It reduced the required data fraction by 7.37 percentage points on average (95% stratified bootstrap interval [5.71, 9.27]), or 11.40% relative to the matched baseline ([5.87, 17.03]%). Rényi-2 lowered the mean threshold on all six tasks, while spectral entropy gave the lowest thresholds on two heterogeneous arithmetic mixtures. On UCI Superconductivity, every ASR objective improved all six mean test metrics under validation-selected and fixed-budget evaluation, although no objective won every metric. Full-run ModAdd trajectories show recurrent spectral-tail contractions aligned with transient accuracy drops. ASR is therefore a measurable representation-level bias whose benefit depends on architecture, task, and spectral objective.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
grokking, representation learning, regularization, spectral regularization, activation spectra, algorithmic generalization, sample efficiency

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bdDWOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a new regularization method called activation-spectrum regularization (ASR) which penalizes the spectrum of hidden-layer activation covariance matrices over mini-batches, with the penalty increasing as the model becomes more confident in its predictions (likelihood-gated pressure). Specifically, paper compares three objectives: a Rényi-2 participation-ratio proxy, spectral entropy, and a ridge Log-Det penalty, all intended to complement conventional weight decay by constraining representations rather than model parameters.

Proposed approach is evaluated on six task/architecture combinations: modular addition, permutation composition, and parity, including both multilayer perceptrons (MLPs) and Transformers. They report that ASR reduces the average training-data fraction required to achieve 99% test accuracy by an absolute 7.37%, or 11.40% relative to the baseline, across the six-tasks. Rényi-2 reduces the mean threshold on all six tasks, while spectral entropy performs particularly well on heterogeneous arithmetic mixtures. An additional experiment on UCI Superconductivity regression reports improvements in mean test R2R^2, mean absolute error, and root mean squared error for all three ASR objectives, without tuning the regularization coefficient on that dataset.

Critical strengths

None

Critical weaknesses
  1. Empirical support for proposed regularizers is on quite small datasets and network sizes. For this reason, practical scalability and transfer to larger datasets and models remain uncertain.
  2. Primary results are for training data efficiency. However, data efficiency -- unless it is orders of magnitude -- is typically not a concern in typical model training scenarios these days.
  3. Missing reference to and comparison with some closely related works:
    • Kazanskii (2026) — How to Tame Grokking: Representation Geometry as a Control Signal
    • Yunis et al. (2024) — Grokking, Rank Minimization and Generalization in Deep Learning
    • Junior et al. (2025) — Grokking Beyond the Euclidean Norm of Model Parameters
  4. The proposed mechanism is not established causally. The observed spectral contractions coincide with temporary accuracy drops, but the paper explicitly acknowledges that this alignment is not causal evidence. It does not establish that activation-spectrum compression causes grokking, that the model discovers a simpler algorithm, or that lower effective dimensionality is the mechanism responsible for improved generalization. The relationship between the regularizer and the learned computation remains an open question.
Questions for authors

Anything the authors can provide to address the critical weaknesses above will be helpful.

Reviewer_rQQFOpenReview ↗
Rating
4(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes activation-spectrum regularization (ASR) to encourage generalization by concentrating the spectrum of centered hidden-state covariance. It compares a participation-ratio proxy for Rényi-2 entropy, spectral entropy, and a normalized ridge Log-Det penalty, with regularization strength controlled by a detached likelihood-dependent multiplier. Experiments examine the training-data fraction required to reach a specified generalization threshold, training time, heterogeneous algorithmic tasks, and transfer to superconductivity regression. The strongest results are substantial reductions in required training-data fractions on modular addition and group composition. Results on Transformers are more dependent on the task and spectral objective.

Critical strengths
  1. The reported improvements are substantial in several controlled settings. In Table 1, spectral entropy reduces the ModAdd data threshold from 66.5% to 49.2%, while Log-Det reduces the S5 threshold from 55.9% to 33.6%. The heterogeneous-task results are also encouraging: Rényi-2 and spectral entropy succeed in all five three-task runs, compared with one of five for the baseline.

  2. The experimental reporting includes matched seeds, explicit accounting for unsuccessful runs, and an aggregate analysis that preserves dependence arising from shared baselines. The regression evaluation uses validation-selected checkpoints and separately reports fixed-budget performance.

  3. The participation-ratio objective provides a computationally attractive implementation without an eigendecomposition. The small-matrix determinant identity is correct, and the authors explicitly report its finite-precision limitations.

  4. The paper appropriately limits its conclusions. It distinguishes spectral concentration from program complexity, acknowledges task-dependent effects and computational overhead, and does not present the spectral trajectories as establishing a unique causal mechanism.

Critical weaknesses

Post-rebuttal update, 15 November 2026: I now recommend 4: Clear acceptance. The specified search and regression protocols, explanation of the baseline discrepancy, corrected trace-normalized gradient, gate ablation, and five-seed MNIST extension sufficiently address my principal concerns for the paper's scoped empirical contribution. The revision appropriately acknowledges that gating has no established uniform speed advantage and that the spectral comparisons do not establish a universal mechanism. The nonmonotone threshold search, exploratory MNIST tuning, and missing matched threshold/weight-spectrum baselines remain limitations. See my follow-up comment. The original concerns below are retained for context.

  1. The distinctive role of likelihood gating is not established experimentally. The paper does not compare the proposed method with the same activation-spectrum penalty without the likelihood factor or with a simple delayed regularization schedule. The additional g(R) multiplier also changes the effective penalty strength. Consequently, the experiments do not isolate whether the improvements arise from activation-space regularization, confidence-dependent feedback, or the penalty-dependent multiplier. A compact ablation on ModAdd and S5 would be informative. A direct comparison with the weight-spectrum entropy regularization discussed in DeMoss et al. would also clarify the contribution.

  2. The principal data-threshold metric is insufficiently specified. Section 3.2 describes a “binary-search-style” procedure without specifying the search resolution, stopping rule, subset construction, or initialization policy. It is unclear whether subsets are nested, whether each candidate fraction is trained from scratch, and whether the held-out set remains fixed. Moreover, success within a finite optimization budget need not be monotone in training-data fraction, which complicates binary search and the interpretation of unsuccessful searches. Pseudocode and direct evaluations around several estimated thresholds would strengthen this evidence.

  3. Important experimental details need clarification. Equations 10–12 define the gate through classification cross-entropy, but the regression experiment does not specify the corresponding gate, training loss, target normalization, or train/validation/test proportions. If the gate uses exp(-MSE), its behavior depends strongly on target scaling. In addition, Table 4 reports 14.0k steps for baseline ModAdd at weight decay 0.1, whereas Table 5 reports 28.8k steps without dropout under apparently matching settings. An undocumented configuration difference may explain this, but the discrepancy needs reconciliation.

  4. The main comparisons fix weight decay at 0.1 despite the strong sensitivity to weight decay shown in Table 4. The reported gains therefore establish an improvement over the selected configuration, but leave uncertainty about the advantage over a comparably tuned baseline.

Questions for authors
  1. How does likelihood-gated ASR compare with the same spectral penalty without the likelihood factor, using comparable coefficient tuning? Does the g(R) multiplier contribute independently?

  2. Can the authors provide the exact data-threshold search algorithm, including subset nesting, initialization, held-out-set construction, search resolution, and treatment of nonmonotone outcomes?

  3. What regression loss, likelihood-gate definition, target preprocessing, and data-split proportions were used? What configuration difference explains the baseline ModAdd results in Tables 4 and 5?

  4. Do the principal threshold improvements persist against a comparably tuned weight-decay baseline and a weight-spectrum entropy baseline?

  5. Could the authors clarify the gradient interpretation in Equations 8–9? Equation 8 differentiates raw Log-Det with respect to covariance eigenvalues, whereas the implemented objective differentiates through trace normalization. The resulting normalization term should be included when interpreting activation-mode updates.

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

The paper analyzes whether a penalty that pushes the hidden activations of a network to use fewer directions (activation-spectrum regularization) helps models to grok with less training data. Concretely, they add three different versions of the penalty on top of standard weight decay and make the penalization stronger as the model becomes more confident on its training data. They then measure the smallest share of training data with which the model is still able to grok within 100k steps, and study this on three tasks. They find that their penalty lowers the required amount of data in most settings.

Critical strengths

The idea of the paper is good, and the execution is done well and is relatively comprehensive (with multiple different regularizations, tasks, and models).

Critical weaknesses

One aspect to investigate more deeply is whether the gains really come only from the spectral penalty, rather than from also generally regularizing more. Weight decay often plays a significant role in grokking, meaning that the combination of these effects should potentially be investigated more.

Questions for authors
  1. Have you looked in more detail into what the correct regularization in terms of weight decay is for the different settings and comparisons, and how weight decay affects your comparisons?
  2. Also, do you have thoughts on how these ideas could be transferred to larger models?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.