acceptodds
Withdrawn from ICLR 2027

Inert on Algebra, Disruptive In-Context: A Causal Ledger for Rank Collapse in Grokking

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Effective-rank collapse is one of the most reproducible structural correlates of grokking. We ask whether preventing it changes generalisation, using a rank pin on the token embedding that is magnitude-matched online to the force driving the natural collapse, and we run the same pin on both sides of the comparison. On modular arithmetic the collapse is dispensable: pins at 0.9 times initialisation leave grokking on time under two pinned rank estimators and under a weight-decay-free optimiser, and still generalise on the attention-only architecture our in-context tasks use. Capping the embedding's Fourier concentration only paces grokking, with a pacing threshold of 0.118 at a 60k-step horizon (profile interval 0.110 to 0.127). On two sibling in-context circuits the same pin has the opposite effect. On bursty in-context labelling (BICL) none of 5 pinned seeds generalises by 1.3M steps, 45 to 97 times each seed's own unpinned grokking step, with floors held; on 1-hop induction the circuit does not form in any of 8 pinned seeds, against 8 of 8 paired baselines (exact paired p = 0.004). Each circuit has its own stable-rank threshold, 0.600 at 200k steps (bootstrap 0.550 to 0.666) and 0.845 at 75k (profile 0.804 to 0.895), and neither pre-registered normalisation carries one to the other, so the threshold is task-local. The block is not carried by the value of the global rank statistic, since matched random pushes keep rank near initialisation and generalise; it tracks how firmly the constraint binds. Four pre-registered mechanism attempts returned negative, so the mechanism remains open. Under standard softmax cross-entropy, weight decay drives the post-memorisation trajectory, and more than a quarter of the collapse happens after generalisation. Finally, the same statistic is timing-blind when read from the input embedding and tracks the grokking step when read from the unembedding, which explains how the correlate was misread.

Primary area:
interpretability and explainable AI
Keywords:
grokking, mechanistic interpretability, delayed generalization, causal intervention, effective rank, in-context learning, induction heads, training dynamics

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

The paper tries to connect rank collapse with the phenomenon of grokking.

Critical strengths

The authors address an important research question, that is trying to understand whether rank collapse is necessary for grokking.

Critical weaknesses

The paper is very badly written and, because of that, very difficult to understand. Many things are not defined properly, the jargon introduced for the first time by the authors is left for granted, without explaining it. In this way, I have barely identified the main topic and question that the authors try to address in their work, but I cannot definitely understand which is the main finding of this paper, nor how their results compare with the existing literature, since a "related works" section is completely missing.

Because of that, I am forced to reject the paper and encourage the authors to resubmit their work after a consistent rewriting of the paper.

Questions for authors

The questions are many since the paper is very hard to read.

For instance, in the abstract, the authors write:

"We ask whether preventing it changes generalisation, using a rank pin on the token embedding that is magnitude-matched online to the force driving the natural collapse, and we run the same pin on both sides of the comparison."

What does "rank pin" means? "magnitude" of what? which "force"? "both sides of comparison"...which comparison are we talking about?

In the introduction, the authors write:

"We answer with a measured contrast, obtained by running one causal intervention on both sides of a dissociation. The intervention is a pin that holds a chosen rank statistic near its initial value, magnitude-matched on each task to the force that drives its natural collapse, and replicated penalty-free by hard projection."

What does "dissociations" means? What "rank statistics" are we talking about? That is, rank of what? "hard projection" of what?

I could keep on going for the whole paper. In this way, it is impossible to understand the author's main results.

Reviewer_wNJyOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work starts by asking whether rank collapse is necessary for grokking during training. By using Participation Ratio (PR) and Stable Rank (SR) as intervention targets and separating “leading” from “timing-predictive” signals, the paper finds a clear contrast across task settings. For modular arithmetic, preventing rank collapse does not prevent grokking, suggesting that the collapse is dispensable. In contrast, on two in-context learning tasks, the same rank-preserving intervention blocks generalization or prevents circuit formation within the evaluated horizons. The work also shows that the same rank statistic plays different temporal roles in the embedding and unembedding matrices: the embedding rank collapses early but is largely timing-blind, whereas the unembedding rank closely tracks the grokking transition.

Critical strengths
  1. This work uses the Participation Ratio(PR) and Stable Rank(SR) as a supervision signal and incorporates them into the loss function. This work uses rank pin to represent it. And this work leaves a great direction for future work, which is why preventing rank collapse would harm the generalized ability of the ICL circuit. Even though this work does not find the root reason, the authors have tried different perspectives to figure it out. So I do not think this should be regarded as a weakness.
  2. This work provides a useful temporal analysis that separates “leading” from “timing-predictive” signals. In particular, it shows that the effective-rank collapse of (WE)(W_E) occurs early but carries little information about the seed-specific grokking time, whereas (WU)(W_U) rank and Fourier concentration closely track the transition. This distinction refines the interpretation of rank collapse as a progress signal.
Critical weaknesses
  1. Although the intervention strength is matched by gradient norm, matching the magnitude of the intervention does not control its direction in parameter space. The rank-pin gradient may therefore alter optimization dynamics or representation geometry beyond the intended rank statistic. This work partially addresses this concern with matched-random-push controls and appropriately avoids claiming that rank itself is causally operative. But the current controls do not fully isolate the effect of the constrained rank statistic from the effect of imposing a persistent structured counter-force.
  2. The empirical scope remains narrow. This work remains unclear whether the observed algebra-versus-in-context dissociation extends to structurally different in-context mechanisms, larger models, or natural-language settings.
Questions for authors
  1. A seed-level illustrative trajectory comparing an unpinned run, a rank-pinned run, and a matched-random-push run on one algebra task and one in-context task would make the causal contrast easier to follow. In particular, jointly plotting validation accuracy, PR/SR, and the circuit-strength measure would help readers see exactly when the intervention becomes active and how the subsequent trajectories diverge.

  2. I suggest this paper could discuss related work more deeply. Because the journey to find the reason is a process. For example, the authors can discuss Li (2026), “Low-Rank Decay for Grokking in Scale-Invariant Transformers.”[1]

    That work reports that explicitly promoting effective-rank collapse through a spectral regularizer can facilitate grokking and expand the generalizing regime in modular-arithmetic Transformers. At first sight, this appears to be in tension with the present finding that preventing rank collapse does not prevent algebraic grokking. However, the two results may instead distinguish facilitation from necessity: rank collapse may provide an efficient optimization route toward generalization without being required for a generalizing solution to exist. A direct discussion of this distinction, as well as the differences in constrained matrices and model assumptions, would significantly strengthen the positioning of the paper.

[1] Li M. Low-rank decay for grokking in scale-invariant transformers: a spectral-geometric view[J]. arXiv preprint arXiv:2606.04405, 2026.

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

The paper studies whether the drop in effective rank observed during grokking is necessary for generalisation. First, the authors study a 1-layer transformer model trained on modular arithmetic. During training, they test the effect of constraining the rank of the embedding matrix by adding a penalty when the rank falls below 90% of its initial value. They find that models trained with that penalty still grok after roughly the same time. The authors also experiment with a regularisation that limits how strongly the embedding concentrates in the five strongest frequency components. This delays generalisation, but the model still grokked when trained for longer.

The authors then apply the embedding-rank penalty to two ICL tasks (bursty in-context labelling and 1-hop induction), using a two-layer attention-only transformer. On both tasks, the constrained models fail to generalise, while the unconstrained models successfully generalise. Interestingly, they find that equally strong random perturbations allow generalisation, while also maintaining a high embedding rank. This suggests that high rank along doesn’t explain the failure, but the authors do not present an explanation of this phenomenon.

Critical strengths
  • The paper successfully demonstrates that rank collapse is not a necessary condition for grokking and learning the two studied ICL tasks in a simplified setting. The comparison with random perturbations is an interesting idea.
Critical weaknesses
  • I am not convinced that testing whether rank collapse is necessary for grokking in this narrow setting addresses a meaningful question about generalisation. I have understood rank collapse primarily as a sign that training removes parts of a memorised solution as a compact algorithm becomes dominant (especially on small, specialised tasks). While the paper provides direct evidence that the measured rank collapse is unnecessary for grokking on its algebraic tasks, which goes beyond this intuition, it does not clearly establish why this was a compelling hypothesis or how refuting it advances our understanding of generalisation.
  • It remains unclear whether preserving rank changes the computation the model learns. One possibility is that the constrained model develops essentially the same Fourier circuit while retaining additional directions that contribute little to its predictions. This would explain how the model maintains high rank without changing its solution. The authors should compare the computations learned with and without the constraint and test whether removing these additional directions affects predictions.
Questions for authors
  • Do the rank-constrained arithmetic models use the same Fourier circuit as the unconstrained models? Can the additional directions maintaining high rank be removed without substantially changing their predictions?
  • What happens if the rank penalty is introduced after a model already generalises?

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

Related papers

Discussion (0)

Sign in to comment.