acceptodds
Under review as a conference paper at ICLR 2027

Misalignment of Low-Loss Regions Causes Grokking

Abstract

Grokking refers to the delayed emergence of validation-set generalization after a model has already overfit the training set. Although first observed in small algorithmic tasks trained with transformers, its underlying mechanism remains unsettled. In this work, we develop an analysis framework based on mode connectivity and the geometry of low-loss regions. The framework predicts that the standard modular-arithmetic setting does not always produce grokking: under a symmetry-preserving train/validation split, we observe a stable anti-grokking case in which validation performance does not recover. This counterexample challenges several existing correlational explanations of grokking. More broadly, our analysis framework and results further suggest that grokking arises when the low-loss regions induced by the training and validation partitions are misaligned. Once these regions become well aligned, training hyperparameters alone cannot produce grokking and the observed dynamics collapse to either trainable or non-trainable behavior.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
grokking, low-loss-region

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_7oxwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes that grokking is a result of train-validation set misalignment between low loss regions (LLR). They further suggest that known indicators of grokking (Fourier circuits, Slingshot effects, etc.) are correlates of this more fundamental cause. Specifically, the model has to move from a train-set-only LLR to a LLR that applies to both train and validation set. Hyperparameter (e.g. weight decay or learning rate) effects are then said to affect the transition from the former to the latter. An additional implication is that if the LLRs are aligned, no grokking occurs.

To support this idea they provide the following support:

  • An empirically supported conjecture that the dataset-wide LLR is a single path-connected region.
  • An "anti-grokking" scenario, where they manipulate the train/valid split in the original Power et al. modular-arithmetic setting to prevent grokking while not changing indicators to an extent.
  • A phase diagram analysis that shows that as the train set increases (through nested subsets of an overall dataset), memorization decreases and comprehension tends to increase across multiple hyperparameter settings.
Critical strengths
  • The overall perspective on loss landscape geometry / mode connectivity is intuitive and easy to follow. Also, they reinterpret existing literature on the subject instead of discarding it.
  • The "anti-grokking" experiment is a genuinely interesting modification to the existing modular-arithmetic of the original grokking paper. The implications of symmetries among splits (and effective information entropy carried by the train set partition) is worth highlighting and probing further.

I name these points to motivate my suggestion for a "weak reject" instead of a strong one. If the critique below can be addressed I can see it warranting an acceptance.

Critical weaknesses
  • The description of grokking is bordering on circular
    • The "analysis framework" is schematic. For example, it does not suggest why or when a model would move from LLR(D_train) to LLR(D_train + D_val). Having to move from LLR(D_train) to LLR(D_train + D_val) is part of the definition of grokking and in the absence of a misalignment it does not need to happen since most of the healthy training outcomes are in LLR(D_train + D_val). Without some description of why the transition would occur it is hard to see it as a satisfying account of the cause of grokking.
    • "Alignment" is never defined precisely or measured quantitively even though it is part of the core thesis of the paper. Without defining alignment before probing it in an experiment like Fig 7 it is difficult to determine if the outcome reflects alignment or something else. In Fig 7 the aim is to suggest alignment but then a lack of alignment (through a lack of effective information carried by the train set) is invoked to justify the lack of grokking.
  • Experimental support is light
    • The empirical support for conjecture 1 is based on an experiment presented in the appendix and it only extends to LLR(D_train + D_val), not to LLR(D_train) which also needs to be traversed by the model. Furthermore, this result is used to justify Fig 6 (which relates to a completely different dataset).
    • The dismissal of slingshot effects or fourier components as indicators of grokking is based on two single qualitative plots per setting (end-of-training Fourier spectra; short early windows of the training-loss curve).
    • The main evidence for alignment (the phase diagrams) are based on rather crude classification criteria (Appendix C). Even then the results show mostly confusion and no actual grokking.
Questions for authors

Request for clarification

  1. Do you have a more precise definition or measure of alignment that can be used as reference from which to interpret the results in Fig 7?
  2. Can you account for why the model would move from LLR(D_train) to LLR(D_train + D_val), even speculation?
  3. Does the epoch scaling in Appendix A ("linearly scales down as the number of training samples increase") apply to the Fig 7 sweep? If so, how many optimizer steps does each n receive, and is the batch size max or min(len(D), 65536)? A much shorter budget at large n could by itself explain the shift from Memorization to Confusion and the absence of Grokking.

Additional experiments

  1. Do you have experiments to show that LLR(D_train) is also a single path-connected region?
  2. Can you provide a more quantitative account of the slingshot effect and fourier spectra in the anti-grokking scenario?
  3. I would expect the results in Fig 7 to show a transition from Memorization to Grokking to Comprehension if the increase in n really makes the LLR(D_train + D_val) more significant. Can you produce such an experiment? Maybe with smaller jumps in weight decay?
3 comments

We thank the reviewer for the careful reading and constructive comments.

The "analysis framework" is schematic. For example, it does not suggest why or when a model would move from LLR(D_train) to LLR(D_train + D_val).

Newly added Appendix J explains why optimization continues after training accuracy reaches 100%. Training accuracy records only whether the prediction is correct; it does not imply that the training cross-entropy or its gradient is zero. Because the modular dataset is noise-free, correctly classified examples can still have nonzero cross-entropy. Therefore, the training loss can continue decreasing after training accuracy reaches 100%.

The model can consequently move through the training-only region and into the overlap when that overlap is accessible. The eventual destination depends on the effective information carried by the training partition, as illustrated by the anti-grokking example. In this sense, the LLR geometry determines which regions are available, while the optimization settings determine the trajectory through them.

"Alignment" is never defined precisely or measured quantitively even though it is part of the core thesis of the paper.

The ideal measure would be the volume of LLR(train+!val), but computing this volume is infeasible in the high-dimensional parameter space.

Appendix K therefore introduces an optimization-based proxy in Figure 17. Starting from the same initialization, we combine normalized training and validation gradients to decrease training loss while increasing validation loss. The resulting train–validation accuracy gap indicates how easily an accessible train-only solution can be found. We emphasize that this is a proxy for misalignment, not a direct measurement of LLR volume. Figure 7 provides a separate indirect probe through the distribution of phase cells.

In Fig 7 the aim is to suggest alignment but then a lack of alignment (through a lack of effective information carried by the train set) is invoked to justify the lack of grokking.

Misalignment is caused by "effective training information is insufficient for the model capacity"

The lack of grokking is caused by "effective training information is insufficient for the task".

When effective training information increases, if the alignment happens earlier than "sufficient for the task", then grokking will never occur. This is the reason grokking is absent in Fig. 7.

Please see Appendix F for details about how to reproduce grokking.

The empirical support for conjecture 1 is based on an experiment presented in the appendix and it only extends to LLR(D_train + D_val), not to LLR(D_train) which also needs to be traversed by the model.

Under the constraint-based heuristic in Eq. 5, a point with low loss on the full dataset is expected to have low loss on the subset of the dataset under the same threshold.

Thus, full-dataset mode connectivity motivates mode connectivity for subset of the dataset.

We also provide empirical results for the mode connectivity on the training partition of dataset xyx^y to address the reviewer's concern, see Figure 9.

The main evidence for alignment (the phase diagrams) are based on rather crude classification criteria (Appendix C)

Appendix C adapts the criteria from Appendix A of existing grokking work [1]. We replace the fixed number of training steps with an adaptive time point, defined as when training accuracy first reaches 95%95\%, so that the criteria remain comparable across datasets.

Even then the results show mostly confusion and no actual grokking.

I would expect the results in Fig 7 to show a transition from Memorization to Grokking to Comprehension if the increase in n really makes the LLR(D_train + D_val) more significant. Can you produce such an experiment? Maybe with smaller jumps in weight decay?

We appreciate this question because it could be a nice after-class exercise. Newly added Appendix F explains why Grokking cells are absent and how to reproduce them.

In brief, grokking requires the task-learnability threshold (anti-grokking example) to be reached before the alignment threshold. In the D_1997 sweep, these thresholds occur in the opposite order, so the intermediate grokking cells are skipped. Sampling more weight-decay values would not change this threshold ordering.

The solution is to reduce the task complexity by changing the fixed modulus from 19971997 to 997997, while keeping the phase-diagram procedure fixed. Grokking cells reappear at n=197n=197, 397397, and 597597.

[1] https://arxiv.org/pdf/2205.10343

Do you have a more precise definition or measure of alignment that can be used as reference from which to interpret the results in Fig 7?

We provide a quantification in Appendix K, Figure 17.

Mathematically, a natural definition of misalignment is the accuracy gap between the training and validation partitions, integrated over the volume of LLR(D_1 + !D_2). However, this region lies in RdR^d, and its volume is infeasible to compute in a high-dimensional parameter space. Therefore, we use an indirect measure instead of attempting to estimate the full volume.

Specifically, we use an optimization procedure to find an accessible point in LLR(D_1 + !D_2) with low training loss and high validation loss. The size of the train–validation accuracy gap, together with how easily such a point can be found, serves as a proxy for misalignment. A larger gap and easier access indicate greater misalignment, whereas a smaller gap or failure to find such a point indicates better alignment.

Can you account for why the model would move from LLR(D_train) to LLR(D_train + D_val), even speculation?

The full answer is provided in Appendix J.

This question can be answered by plotting the grokking training-loss curve on a logarithmic scale. The plot shows that training loss continues to decrease after training accuracy reaches 100%. Because the modular dataset is noise-free, correctly classified examples can still have nonzero cross-entropy. Therefore, the training loss can continue decreasing and its gradient can remain nonzero after accuracy saturates. The eventual destination depends on the effective information carried by the training partition, as illustrated by our anti-grokking experiment.

Does the epoch scaling in Appendix A ("linearly scales down as the number of training samples increase") apply to the Fig 7 sweep? If so, how many optimizer steps does each n receive, and is the batch size max or min(len(D), 65536)? A much shorter budget at large n could by itself explain the shift from Memorization to Confusion and the absence of Grokking.

The epoch scaling also applies to the Fig 7 sweeps. The scaling ensures that the number of optimization steps is greater than the n=97 baseline as the task becomes more complicated. We add the formula to Appendix A.

The total batch size and epochs are listed in the table below.

nnBatch sizeEpochs
979,409150,000
19738,809150,000
29765,536150,000
39765,53653,034
49765,53653,034
59765,53628,868
69765,53618,750
79765,53613,417
89765,5368,100
99765,5366,630
199765,5362,000

Can you provide a more quantitative account of the slingshot effect and fourier spectra in the anti-grokking scenario?

Reviewer 1 correctly pointed out that our result does not refute the Slingshot claim: it shows that Slingshots can occur without grokking by changing dataset types, so their presence is not sufficient to guarantee grokking. We have revised the text to state this distinction precisely.

Regarding quantification, the original Slingshot paper likewise characterizes the effect using training loss and accuracy rather than a separate quantitative metric. So we do not have quantitative metric to compare to, either.

Reviewer_7oxwOpenReview ↗

I don't believe the work can be published at this point

Hi authors!

  • (Appendix J) I understand that the training loss can keep decreasing after 100% train accuracy. Appendix J explains why optimization continues, but not why it ends up in the overlap LLR(D_train + D_val) rather than staying in the train-only region LLR(D_train + !D_val). That part is deferred to "effective information", which is not defined. Without it, the account of why grokking happens remains a description of the phenomenon rather than an explanation of it.
  • (Appendix K) My request for a measure of misalignment was not for its own sake. The proxy in Appendix K is only reported; it is not used to support the claims in Fig. 7, Appendix F, or the anti-grokking split. I asked for it so that misalignment could be established independently of the outcome. Currently, the explanation for the absence of grokking in Fig. 7 (the regions align before the task becomes learnable) is inferred from that absence itself, since neither threshold is measured. This is the circularity I am concerned about.
  • (Appendix F) This is progress towards moving away from a schematic account of grokking, and more than "a nice after-class exercise". The ideas of effective training information entropy, task-relevant information, task complexity and model capacity are in the right direction for characterizing why grokking happens. However, as the appendix itself states, these are "qualitative bookkeeping variables rather than formal measurements", which is too imprecise for a principled characterization at this stage.

I suggest developing the ideas in Appendix F further. I feel the work cannot be published in its current form.

Reviewer_XLiAOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors explore grokking through the insight that grokking represents a trajectory through parameter space. They find that, for a specific adversarial dataset selection for the classic modular arithmetic grokking setup, they still see previously identified hallmarks of grokking without improved performance on the validation set. They also show that, as low loss regions shrink and are more aligned, grokking disappears as there is no longer a training trajectory that passes through the "success only on training data" region.

EDIT: Score updated to 2) weak reject

Critical strengths
  • Figure 3 (the parameter space diagram) is a great intuition pump, and framing grokking in terms of parameter-space vectors is insightful.
  • Specifically designing datasets to understand how grokking corresponds to trajectories in parameter-space is a great idea for experiments.
Critical weaknesses
  • One of the main claim of this paper seems a bit muddled: is it that previous results about grokking (slingshot effects, Fourier features, etc.) are not sufficient? That they're not necessary? I don't think that these previous explanations are mere correlates with grokking; for modular arithmetic, my intuition is that the development of Fourier features is necessary but not sufficient.
  • It is somewhat surprising that partitioning the modular arithmetic datasets such that all symmetric pairs are in either the training or the validation set. Still, it's not clear to me that this invalidates other grokking research? For example, we wouldn't expect training models on some subset of modular arithmetic to result in grokking on, say, completion of Bulgarian folk song lyrics. But we would still probably see the model's solution on modular arithmetic iterate through various degrees of generalization during the training process. So it seems like the real question is "to what datasets is it possible for the current dataset to generalize"?
  • While still insightful, the claims that grokking is not (insert previous explanation) but is, instead, parameter-space misalignment are almost tautological. By definition, grokking must involve passing through a solution that only works on the training set but not the validation set into a region that works on both.
Minor Suggestions/Clarity
  • Section 2.1: Using variations of D for both the number of trainable parameters and the dataset makes the notation a bit hard to follow
  • Figure 3: I appreciate the reasons to write defensively so people don't misinterpret the figure, but maybe put the clarifying statements about what the figure is not in a footnote?
  • Figure 4(a) makes the dataset partition more confusing, I feel. A table with a few examples might be more clear?
Questions for authors
  • The authors mention singular learning theory, but, other than discussing loss landscapes, the connection is not clear. Can you elaborate?
  • The explanation of shared information between the training and validation sets is underdeveloped — I'm actually curious what you think here?
  • From Figure 3, I thought you were going to do something like train models on the train set, train other models on the validation set, and interpolate between them to characterize the solutions in parameter space. This isn't exactly a question, but something like this would be interesting.
  • What else would this parameter-space view of grokking predict?
4 comments

We thank the reviewers for their time and effort on our paper.

I don't think that these previous explanations are mere correlates with grokking; for modular arithmetic, my intuition is that the development of Fourier features is necessary but not sufficient.

We agree that Fourier features may be necessary for grokking. However, the Fourier work proposes to use the Fourier circuit as a progress measure, which is a much stronger claim. Our results show that the circuit can develop without any validation-set generalization.

Regarding Slingshot, Reviewer 1 points out that the Slingshot paper does not guarantee grokking with Slingshot. We have revised our paper to present this point more clearly.

For example, we wouldn't expect training models on some subset of modular arithmetic to result in grokking on, say, completion of Bulgarian folk song lyrics. But we would still probably see the model's solution on modular arithmetic iterate through various degrees of generalization during the training process. So it seems like the real question is "to what datasets is it possible for the current dataset to generalize"?

This comparison changes the task and is outside the scope of our paper. We make no claim that a modular-arithmetic training set should generalize to Bulgarian folk lyrics. In the anti-grokking experiment, the task, model, and complete modular-arithmetic dataset are fixed; only the train/validation partition changes.

While still insightful, the claims that grokking is not (insert previous explanation) but is, instead, parameter-space misalignment are almost tautological.

This criticism conflates a necessary condition with our main contribution. We agree the statement "grokking requires a trajectory from a training-only low-loss region to an overlapping low-loss region" is tautological.

However, this is not our main contribution. Our contribution is to use this geometry as a framework for predicting when such regions are aligned or misaligned, and for separating the factors that determine the geometry from those that determine the trajectory.

Sections 4 and 5 argue that the dataset and model architecture primarily determine low-loss-region alignment, whereas learning rate, weight decay, initialization, and related choices primarily determine the training trajectory. This framework predicts that hyperparameter changes alone cannot create grokking when the low-loss regions are already well aligned.

In addition, we encourage the reviewer to read Appendix F, where we reproduce Grokking in the setting of Figure 7. The explanation is not tautological: the fact that Reviewer 3 was unable to answer this question correctly shows that the distinction is not immediate.

Section 2.1: Using variations of D for both the number of trainable parameters and the dataset makes the notation a bit hard to follow

We revised Section 2.1 so that dd denotes the number of trainable parameters

Figure 3: I appreciate the reasons to write defensively so people don't misinterpret the figure, but maybe put the clarifying statements about what the figure is not in a footnote?

We have moved the explanation to footnote.

Figure 4(a) makes the dataset partition more confusing, I feel. A table with a few examples might be more clear?

Figure 4(a) is a simplified version of Figure 1(a). They share the identical color scheme except that Figure 1(a) also shows the label, while Figure 4(a) removes it. In addition, Figure 4(a) shows a clear symmetry relationship across f(x)=x, use a table cannot clearly present this relationship.

The authors mention singular learning theory, but, other than discussing loss landscapes, the connection is not clear. Can you elaborate?

SLT motivates our view of neural-network solutions as extended low-loss valleys with low-loss directions near solutions, rather than isolated minima. Mode Connectivity provides complementary empirical evidence that independently trained solutions can be connected by low-loss paths.

However, mode connectivity lacks theoretical evidence and remains an empirical observation. Thus, we use both SLT and MC to support Conjecture 1. This is also the reason why we say "There is also theoretical support for this perspective, please see Appendix O".

[1] https://arxiv.org/pdf/2301.05217

The explanation of shared information between the training and validation sets is underdeveloped — I'm actually curious what you think here?

Our paper never mentions "shared information between the training and validation sets". The reviewer probably conflates it with the “overlapping low-loss region of the training and validation partitions.” These are distinct concepts: shared information is defined in the input/task space, whereas an overlapping low-loss region is defined in the model-weight space.

For the overlapped low-loss region, please refer to (Line 96, initial submission) for definition and (Line 221, initial submission) for the discussion of whether the overlapped low-loss region is empty or not.

What else would this parameter-space view of grokking predict?

Our paper makes several important claims in Section 5. These points are not reflected in the review, so we restate them below.

Grokking is governed by two levels of factors. The primary factors—the model architecture and dataset—determine the alignment of the low-loss regions. The secondary factors—weight decay, learning rate, initialization, and related choices—control the training trajectory (Line 452, initial submission).

Section 5 also states that grokking is a warning sign that the training dataset is insufficient for the model architecture (Line 470, initial submission).

In the revision, we added two paragraphs to Section 5 explaining that our framework is generally applicable and can help interpret other studies of grokking.

In addition, we encourage the reviewer to read Appendix F, where we reproduce grokking for the nested modular-dataset setting used in Figure 7.

Rather than getting caught in the weeds, I'd like to offer some high-level constructive feedback, since I like the general idea of the paper and I do think the authors are onto something promising (even though it's not there yet). My interpretation of what the contribution is/could be:

  1. If we view grokking as a trajectory in the loss landscape in parameter space, we can think of it as passing through misaligned regions, where the model performs well on the training set but poorly on the test set, before entering a region where the model performs well on both.
  2. The loss landscape is defined by the model and the dataset, whereas the training trajectory is a function of hyperparameters, etc. Previous explanations confuse these a bit, so here's why this is a clarifying framework.
  3. We have some way to show/measure/predict when regions in the loss landscape will be misaligned, either as some function of the datasets or by analyzing the loss landscape geometry directly, and we can reliably show grokking in these settings, some of which are not obvious without our insights. When we do not have these misaligned regions, we do not see grokking, despite trying kind of hard to find it. Some of these examples are also not obvious a priori without our insights. Or, we do some sort of principled parameter space inteprolation to find regions of grokking or generalization, etc. that align with our predictions or empirical training trajectories.

The point about Bulgarian folk songs was more of a motivating example than a specific claim. Clearly, we do not expect training on modular arithmetic to transfer to Bulgarian folk song lyrics. But, it's not clear what training on modular arithmetic should generalize to. It is a surprising finding that this does not include flipping the order of the variables. Relatedly, my question about "shared information" was about the idea of "effective information entropy": it should account for task knowledge that makes some samples informative about others.

My point is that this seems like a promising direction for 3) above. The addition of Appendix F is also a strong step in this direction (which seems like it should be promoted to the main body of the paper).

I'm happy to raise my score to a 2, but I think this needs more time to cook and a material rewrite.

We thank the reviewer for the summary of our contribution, which accurately captures the main purpose of our paper.

However, there is one point that I may be misunderstanding:

But, it's not clear what training on modular arithmetic should generalize to.

In the modular arithmetic setting, generalization means correctly predicting the outputs for input pairs that were not seen during training. Once the output and input belong to the same task.

Specifically, we construct a dataset using (x+y) mod 97(x+y) \bmod 97 with x,y ranging from 0 to 96.This gives 97×9797\times97 samples, each with input (x,y)(x,y) and output (x+y) mod 97(x+y)\bmod97. We then split these examples into training and test sets. The intended question is whether a model trained on one subset can correctly predict the outputs for the held-out input pairs under the same modular arithmetic rule.

We are therefore evaluating generalization to unseen examples of the same task. We do not expect the model to generalize to (x+y) mod 98(x+y) \bmod 98 or to any other datasets, which would define a different task.

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

This paper studies the phenomenon of "grokking", where neural networks trained on algorithmic tasks first memorize their training set but eventually generalize late in training. The paper first proposes a formalism/language/framework for talking about grokking in terms of loss landscapes, in particular whether the regions of parameter space where the network achieves good train set loss vs. the regions where the network achieves good test set loss overlap. It then goes on to perform some experiments which suggest that some other explanations of grokking are wrong, such as the slingshot mechanism. My favorite part of the paper was when the authors construct train/test partitions where samples (a, b) and (b, a) are always included in the same partition, and find that it is much harder for networks to generalize when trained on train sets with this property than when partitions are chosen randomly. This also means that the notion of a "critical training set size" is somewhat ill-defined.

Critical strengths
  • First, I'd like to sincerely thank the authors for clearly writing the paper themselves instead of primarily using AI. I found the paper to be well-written and understandable.
  • I like your experiment where instead of choosing the train set partition randomly, you always place inputs (a, b) and (b, a) together in the same partition, which prevents the model from generalizing just via the rule that the function it is fitting is symmetric (a commutative operation). This was something I'd always wondered about but had never tested before, and it's interesting that it prevents generalization in the setting you study. You make a good point that the "critical data size" is then somewhat poorly defined.
Critical weaknesses
  • I liked the spirit of your discussion of low loss regions, your Figure 3, and the more visual story that comes along with it. That said, I'm not sure that all this discussion of low loss regions actually advances my understanding about the causes of grokking. When a network memorizes but doesn't generalize, then it's trivially in a part of the loss landscape where the train loss is low! When it generalizes, then it's trivially in the part of the loss landscape where the test loss is also low! Ultimately the thing to be understood with grokking is why for some tasks/networks training first moves to a part of the landscape that you call LLR(M, D1 + !D2) and then eventually moves into LLR(M, D1 + D2). The statement that training proceeds through these different regions is equivalent to just describing the phenomenon of grokking that we already observe from the loss curves. My understanding is that you don't actually provide a new explanation for why the loss landscape regions have the structure they do and why training proceeds through them in a certain order? If not, then a lot of the paper, which the paper is named after, just describes the basic phenomenon of grokking but with slightly different language. This might be an okay contribution, but I'd be curious what the other reviewers think. I'd overall say that it's not a meaningful enough contribution to significantly contribute to an acceptance decision.
  • Your main empirical results are that (1) when you choose partitions so that (a, b) is always in the same partition as (b, a), then your networks don't generalize, which you refer to as "anti-grokking" and (2) that even in this anti-grokking setting, you still observe slingshots in the train loss curves, which may invalidate the slingshot mechanism as an explanation of grokking.
    • I think that (1) is a good observation. Although the claim that it has a very strong effect on generalization (perhaps justifying giving it a dedicated name "anti-grokking") doesn't yet have super strong evidence. Since weight decay is very helpful for networks to generalize in this setting, I would have liked to see some experiments training with weight decay. I wouldn't be surprised if weight decay allows networks to generalize even when you mess with the train partition in this way. So it would only be "anti-grokking" when training without weight decay. Could you run these experiments?
    • I don't think that (2) actually challenges the slingshot mechanism claim as stated by Thilak et al. Those authors claim that, in the grokking setting without explicit regularization, the absence of slingshots implies the absence of generalization, not that the presence of slingshots implies the presence of generalization.
Questions for authors

See my questions above.

2 comments

We thank the reviewers for their time and effort on our paper.

Summary:

We would like to clarify that Sections 4 and 5 contain the paper's main contributions, although they are not reflected in the review summary. Section 3, which presents the anti-grokking example, is also important because it provides the context needed to understand Figure 7. But Section 3 itself is not our major contribution.

Our main result is the analysis of alignment between the training and validation low-loss regions (LLRs), together with the conclusion that the model architecture and dataset are the primary determinants of whether grokking is possible.

Figure 7 studies a controlled, nested family of datasets constructed from Dmod1997\mathcal{D}_{\mathrm{mod1997}}. Adding data adds constraints, so the corresponding LLRs are expected to shrink and become more aligned. Consistent with this prediction, Memorization cells disappear, Comprehension cells emerge, and Confusion becomes more common. The trajectories therefore increasingly fall into trainable (Comprehension) or non-trainable (Confusion).

Section 5 separates two levels of factors. The primary factors—the dataset and model architecture—determine the alignment of the training and validation LLRs. The secondary factors—learning rate, weight decay, initialization, and related hyperparameters—control the training trajectory. When the primary factors produce well-aligned LLRs, changing hyperparameters alone cannot create grokking or memorization. We therefore interpret grokking as a warning that the training information is insufficient for the model architecture.

Thus, the paper addresses the broader question of when the training and validation LLRs are aligned or misaligned. The absence of Grokking cells in Figure 7 is intentional, not an omission. Appendix F explains how Grokking cells can be recovered with the knowledge introduced in our paper. (We do encourage the reviewer to use Appendix F as a check on their understanding of the paper as a whole.)

Weakness:

My understanding is that you do not actually provide a new explanation for why the loss-landscape regions have the structure they do and why training proceeds through them in a certain order.

Our paper do proivde a new explanation around the low-loss regions (a.k.a. loss-landscape regions). See the "misalignment discussion" in Section 4 and Section 5.

Answering why training proceeds through them in a certain order.

Newly added Appendix J explains why optimization continues after training accuracy reaches 100%. Because the modular dataset is noise-free, correctly classified examples can still have nonzero cross-entropy. Therefore, the training loss and gradients can continue decreasing after accuracy saturates. The eventual destination depends on the effective information carried by the training partition, as illustrated by our anti-grokking experiment.

I wouldn't be surprised if weight decay allows networks to generalize even when you mess with the train partition in this way.

We agree that this is a relevant question, but our framework addresses a broader question than whether weight decay alone can recover generalization. It answers how the train/validation partition changes the alignment of the corresponding low-loss regions. In the symmetry-preserving split, pairing (a,b) with (b,a) reduces the effective information in the training partition (Section 4.2). Our framework therefore predicts greater misalignment, more Memorization/Grokking cells, and fewer comprehension cells than for a random partition.

We conducted the corresponding phase-diagram experiment for the anti-grokking split and provide the result here: https://anonymous.4open.science/r/xxxxxxxx-C226/phase_diagram_of_antigrokking/x%2By_mod_97_anti_grokking.pdf. Comparing it with Figure 18, which uses a random partition, supports this broader prediction.

Back to your weight decay question, I believe the answer is already obvious. Note that weight decay is only one secondary factor that influences the training trajectory, it does not change alignment of low-loss regions (Line 452, initial submission).

I don't think that (2) actually challenges the slingshot mechanism claim as stated by Thilak et al.

We agree with the reviewer. After rechecking the Slingshot paper, we find that it states: "Grokking as reported in [17] almost exclusively happens at the onset of Slingshots, and is absent without it." The reviewer is therefore correct that the Slingshot authors do not claim that grokking always happens at the onset of slingshots.

We have revised the paper to state this more precisely.

Reviewer_rNjTOpenReview ↗

Keeping my score

Hello authors. Thank you for your response.

You use "effective information" in your response and in the paper but I don't know what you mean by this.

I feel that I understand the paper as well as I could have reasonably been expected to. I think there are some good elements here, but think that the paper requires a bit more work. I agree broadly with the final response of reviewer XLiA.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.