Where Grokking Happens: Distributed Utility and Fourier Recoding Without a Module Switch
cs.LG · cs.AI
Quick verdict
Grokking as distributed Fourier recoding, not a module switch: rigorous frozen exact-game evidence in tiny models; partly utility-sensitive, limited generality.
Summary
The paper asks where in a Transformer the memorization-to-generalization transition of grokking is functionally expressed. It introduces Transition Games: fully enumerated, exact Shapley/Owen activation games (players: embedding, attention, and MLP activations; utility: clipped predictive compression) evaluated at behavior-matched pre- and post-transition checkpoints, with seed-, split-, and duration-matched non-generalizing control runs. In two-layer mod-47 addition and division Transformers, all paths gain OOD utility (~0.85-1.01 bits/example), yet block-0 attention shows a relative excess over its neighboring MLP (~.16-.19 bits), prospectively confirmed on disjoint seed cohorts. The proposed 'MLP memorizes, attention generalizes' handoff reverses (-.3314 bits at the memory anchor; 0/12). Owen refinement over degree-two Fourier modes shows selected modes carry 67-92% of the addition contrast, transfer to division in multiplicative coordinates, and their effect flows mainly through block-1 MLP in an exact restoration game (12/12 pairs). A fresh prime-by-width-by-head factorial supports a predeclared head-dimension diagonal at p=43 but not p=53. Limitations include tiny tied-embedding models, mechanism restricted to mod-47 arithmetic, self-administered 'frozen' protocols rather than external preregistration, partial utility-sensitivity of several results, and a weakly powered division confirmation.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- During grokking, OOD utility gains are distributed across embedding, attention, and MLP paths, with a replicable relative excess for block-0 attention (~.16-.19 bits/OOD example) over its neighboring MLP.
- The 'MLP memorizes, attention generalizes' handoff prediction reverses in direction: attention already exceeds MLP in training utility at the memory anchor (-.3314 bits/example; 0/12 pairs).
- Selected degree-two Fourier modes carry 67-92% of the addition transition contrast, are more decodable on held-out inputs, and transfer to division in multiplicative-group coordinates.
- In an exact four-path restoration game, block-1 MLP mediates more of the selected-mode effect than the other three tested paths combined (12/12 disjoint pairs; 2.6440 vs .1273 bits).
- A predeclared head-dimension architecture diagonal is supported at modulus 43 but not 53, rejecting a universal architecture ridge without identifying the moderator.
Strengths
Novel behavior-anchored causal estimand
Transition Games compare the last memory-dominant vs first persistent-generalization checkpoints with paired non-generalizing controls; full coalition enumeration (8 to 512 coalitions) plus efficiency audits below 1e-8 removes sampling and linearization error that limit one-at-a-time ablations and attribution patching (Sec. 3.2, App. N).
Genuine prospective discipline
The block-0 target was selected exploratorily and then tested on disjoint addition (seeds 2301-2312) and division (2401-2412) cohorts with pre-frozen Holm families, exact sign-flip tests, and frozen direction-count rules; behavioral gates (12/12, 11/12) keep failed seeds in denominators and forbid gate-failure rescue (Sec. 4.3, Table 5, App. D).
Convergent multi-method evidence for spectral recoding
Unusually transparent negative results and self-audits
Reproducibility infrastructure
Weaknesses
Very narrow model and task scope
Self-frozen protocols and reuse of trained models
Division replication is underpowered and metric-fragile
Mediation claim is conditional and partly post hoc
Headline reversal targets a single prior account
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is grokking and what does this paper conclude about where it happens?
Grokking is the phenomenon where a neural network first memorizes its training data and only much later generalizes to unseen inputs. This paper argues, using exact activation-game interventions on small Transformers trained on modular arithmetic, that the transition is expressed as a distributed recoding of an existing circuit toward Fourier-like representations — with a relative early-attention (block-0) bias and downstream mediation through the final MLP — rather than as a single module switching roles.
Do the results show that MLPs memorize while attention generalizes?
No. The paper tests that specific prediction and finds the opposite direction: at the checkpoint where the model has fully memorized the training set, attention already shows more training utility than the MLP (about -0.33 bits/example in the MLP-minus-attention contrast, with 0 of 12 runs in the predicted direction). The authors treat this as a reversal of the proposed module-handoff account.
Is the study preregistered and reproducible?
Protocols and analysis code were 'frozen' in version control with tags before the named outcomes were computed, which is auditable but weaker than external preregistration with an independent custodian. Code, compact reports, and checksum-addressed data assets are public on GitHub, and the paper reports byte-identical deterministic replays and per-study SHA-256 archive matches.
Do these findings apply to large language models?
Not directly. All mechanistic results come from two-layer, tied-embedding Transformers trained on mod-47 addition and division, with a factorial spanning only two primes and a 2x2 architecture grid. The authors explicitly caution against generalizing synthetic-task findings to language models without direct evidence.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures and several tables were available only as extracted text and captions; visual trends (e.g., Fig. 1 trajectories, Fig. 4 cell patterns) could not be independently inspected.
- The extraction appears complete through Appendix R, but appendix numbers could not be cross-verified against the stated public release assets or code.
- References dated 2025-2026 (e.g., Hidajat et al., ExPLAIND) could not be checked for accuracy of characterization.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.