acceptodds
Under review as a conference paper at ICLR 2027

Implicit Federation Bias: A Study of Emergent Generalization in Federated Learning

Abstract

Federated learning (FL) changes the optimization dynamics by dividing training across local client trajectories and periodic aggregation, yet when and how this affects the emergence of generalization is unknown. We examine this through grokking, a delayed generalization phenomenon in which models generalize only after a prolonged period of memorization, but whose dynamics have been studied almost entirely under centralized optimization, leaving their behavior under FL unclear. We systematically study this using multiple established grokking setups under FL and find that emergent generalization persists, but its timing and behavior can change substantially. Increasing the number of clients or local optimization steps often delays generalization, while alternative federated optimizers can accelerate it by up to an order of magnitude. Stronger federation effects can produce long plateaus at sub-perfect training accuracy or, in extreme cases, no clear memorization period. In both cases, training and test performance improve together only after a long delay. To explain these effects, we analyze how local functional changes survive aggregation. Local updates are dominated by client-specific features, while aggregation preferentially suppresses these components and retains a greater fraction of globally consistent features; the accumulation of this shared global signal is associated with the onset of generalization. We characterize this as an implicit federation bias over learned structure that can delay, accelerate, or qualitatively reshape emergent generalization.

Primary area:
learning theory
Keywords:
Federated Learning, Grokking, Optimization Dynamics, Federated Optimization, Distributed Learning, Delayed Generalization, Emergent Generalization

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

This paper investigates how federated learning affects the emergence of generalization in grokking settings. Using five established setups, the authors systematically examine the effects of client numbers, local training epochs, participation rates, data heterogeneity, and federated optimizers on generalization dynamics. They further propose Implicit Federation Bias, suggesting that aggregation preferentially suppresses client-specific structures while preserving globally consistent ones, thereby influencing the timing and emergence of generalization.

Critical strengths
  1. The paper offers an interesting perspective by shifting attention from final generalization performance to how generalization emerges during federated training. In particular, it shows that federation can alter the classical grokking dynamics, with training and test performance sometimes improving together after a prolonged plateau.
  2. The authors conduct a systematic experimental study across multiple grokking setups and federated configurations. The functional-level analysis also provides a potentially useful perspective on how aggregation influences generalization.
Critical weaknesses

My main concern is the novelty of the proposed Implicit Federation Bias. Existing studies, such as SCAFFOLD (Karimireddy et al., 2020), have already investigated client drift and the effects of inconsistent local updates on global optimization. While this paper provides an interesting functional-level interpretation, I feel that its additional insights beyond existing federated optimization theories could be more clearly established.

Questions for authors
  1. Could the authors further clarify how Implicit Federation Bias differs from existing theories of client drift and local-global consistency? What additional insights does it provide?
  2. Could these findings offer practical guidance for designing federated optimization algorithms? For example, could functional-component retention be used to improve client updates or server aggregation?
2 comments

Dear Reviewer WMeJ,

Thank you for your review and for your comments on how the work shifts attention from final generalization performance to how generalization emerges during federated training and that we conduct a systematic experimental study, providing a potentially useful perspective functional-level analysis. Below we address your specific concerns and raised questions.

Novelty of findings

We acknowledge that we perhaps did not phrase parts sufficiently well leading to confusion over what is novel and as such have made the changes in the updated version.

Whilst we agree that client drift is well studied and that SCAFFOLD corrects it to improve convergence, our contribution is complementary but addresses a different question: rather than examining how client drift affects optimization convergence, we investigate which functional structures survive aggregation, and how their accumulation changes the emergence of generalization.

Our results motivate a mechanism that we call implicit federation bias, which explains that federation itself can act as a regulariser, showing behaviors such joint training-test accuracy growth, or multiple training plateaus. The main idea is that repeated local training and aggregation creates selection preference over what the global model learns. We understand that aggregation cancelling what clients disagree on is expected, what we show is how specific components contributing to memorisation and generalisation develop during training, and how different levels and factors of federation affect them.

Drift studies tend to treat the client update as one vector and do not make this distinction. We also find that drift and disagreement do not sufficiently predict when generalisation emerges and that most of the acceleration does not come from correcting drift. (Appendix E).

We have also added an additional section linking and differentiating our work from existing grokking literature (see Section 5, paragraph “Relation to other grokking explanations”).

Usage in federated optimization algorithms

We have added a causal intervention (Section 5, Figure 10, Appendix A.10). Amplifying, or keeping only, the shared component GG retained by aggregation reaches generalization 1.9–3.1× sooner in Setup A. This confirms that GG points towards the generalizing solution and through it we can alter the federated optimizer by boosting the learning signal on shared generalisable (or conversely memorised) components to change the learning dynamics such as to reproduce the shown joint training-test increases, to speed up convergence and produce grokking-like behaviour whilst keeping weight norms low.

However, because GG relies on the task’s symmetry, this is currently diagnostic rather than for practical use; we discuss initial work and directions on GG estimation to then use in optimizers.

We would like to thank the reviewer since their feedback has helped produce some additional interesting findings.

If this has addressed your concerns, we’d be grateful if you could revisit your score. We also welcome additional questions and discussion as needed.

— The Authors

Thank you very much for your detailed response and thoughtful clarifications. I sincerely appreciate the authors’ efforts in addressing my concerns. Given my relatively limited expertise in this specific area, and after carefully considering the perspectives and feedback from other reviewers, I would prefer to maintain my current rating while deferring to their expertise in the final assessment.

Reviewer_sy84OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates emergent generalization, a phenomenon primarily studied in centralized learning, in the context of federated learning and examines how its dynamics differ between the two settings. Experimental results show that emergent generalization tends to occur later in federated learning than in centralized learning. The authors also identify cases where the post-memorization grokking regime disappears, with training and test accuracy improving almost simultaneously. The paper further examines how various factors, including the number of clients, local epochs, client participation ratio, data distribution, and federated learning algorithms, influence emergent generalization. Finally, the authors introduce the concept of implicit federation bias to explain how federation alters generalization dynamics, suggesting that only certain globally relevant features learned during local training are preserved through model aggregation.

Critical strengths
  1. Interesting observations on generalization dynamics. The paper provides interesting observations on how generalization dynamics differ between federated and centralized learning. In particular, the finding that the separation between memorization and generalization becomes less pronounced as the degree of federation increases is noteworthy.

  2. Analysis across diverse federated learning configurations. The paper provides an interesting empirical analysis of how various federated learning configurations, including the number of clients, local epochs, and optimization algorithms, affect generalization dynamics.

  3. Investigation into the underlying mechanisms. In Section 5, the authors attempt to explain how federation alters generalization dynamics by examining implicit federation bias and the role of global learning rate steps.

Critical weaknesses
  1. Limited novelty of the insights. It is not particularly surprising that generalization dynamics differ between federated and centralized learning. Furthermore, the analyses, including the discussion of implicit federation bias, do not appear to provide new insights beyond these observations.

  2. Confounding effect of the number of data points per client. Although increasing the number of clients strengthens the degree of federation, it also reduces the number of training samples available to each client when the total dataset size is fixed. The latter effect may play a more dominant role in the observed generalization dynamics. It would be valuable to compare these findings with the grokking behavior observed in centralized learning when the training dataset size is reduced, to better distinguish the effects of federation from those of limited local data.

  3. Focus on convergence speed rather than emergent generalization. Most of the analyses and insights appear to focus on convergence speed, particularly how quickly the model reaches a certain level of generalization, rather than on the underlying phenomenon of emergent generalization itself. In particular, many figures primarily measure the time required for test accuracy to reach a predefined threshold, which may not sufficiently characterize the dynamics of emergent generalization.

Questions for authors
  1. Distinguishing federation effects from limited training data. How do the generalization dynamics observed in federated learning differ from those in centralized learning with a similarly reduced training dataset size?

  2. Extraction of individual components. In Section 5, how are the individual components identified and extracted from changes in model outputs?

  3. Inconsistent metrics in Figure 9. Why does Figure 9(a) report the delay as a function of the learning rate, while Figure 9(b) reports only the speedup relative to FedAvg?

  4. Comparison of gradient steps. How are gradient steps accounted for and compared between centralized and federated learning to ensure a fair comparison of their generalization dynamics?

3 comments

Dear Reviewer sy84,

Thank you for your review and for highlighting our interesting observations on generalization dynamics, and investigation into the underlying mechanism. Below we address your concerns and questions.

Novelty

We acknowledge that we perhaps did not phrase parts sufficiently well, leading to confusion over novelty, and, as such, have made the changes in the updated version.

To our knowledge, prior work has not studied emergent generalisation and grokking under federated optimization. While we agree that showing a difference in learning dynamics under these two paradigms would not be surprising, our contribution also addresses a cause behind why federated and centralized dynamics differ. We understand that aggregation canceling what clients disagree on is expected, what we show is how specific components contributing to memorisation and generalisation develop during training, and how different levels and factors of federation affect them.

The obvious explanation of client drift fails to explain what we observe. For instance, intuition from prior work suggests that the further clients move apart, the later the global model generalises, but this wasn’t apparent in our results, with clients’ updates being 9x more aligned and yet generalizing much later (App. E).

We have added an additional section differentiating our work from existing grokking literature (see Section 5 paragraph “Relation to other grokking explanations”).

Federation versus training data effects

Thank you for the feedback; we ran controls as suggested (App. D.3).

First we trained a single model using 1/K of the original training dataset (i.e. a single client partition), sweeping K (Figure 25). As expected the results show that in almost all cases, when reducing the dataset the model fails to generalise within the training budget.

This does not isolate the effects of federation from reduced dataset size as the federated model sees K times more data in total and benefits from other dynamics. In a follow-up experiment, where we fix the data per client so that each client’s share no longer shrinks with K, we find that given total data is the same, whether shrinking the partition by K or fixing the partition size, generalisation speed is not noticeably affected. If data starvation was the cause of generalisation delays then the delay should disappear when the data amount is fixed. The delay therefore comes from splitting optimization across more clients, not merely from the amount of data each client holds.

Focus on convergence speed

We focussed on convergence speed due to its practical implications for training times, however the submitted version already has other metrics and analyses. In particular, we also decompose what is retained in the learning signal when local updates are aggregated by the server, explicitly tracking functional components that contribute to generalization and those that contribute to memorization. We have also now added weight norm details and related findings to existing grokking explanations (Section 5, “Relation to other grokking explanations” paragraph; App. D).

Extraction of individual components

From a saved global state, we replay one round and record how each client's model and the averaged model change their logits over every input pair and every output class. We split each change into four parts by averaging over the algorithmic task's symmetry: operand-only terms, pair lookup, a per-sum table, and the global rule G, which depends only on (a+b−c) mod p. The parts are orthogonal and add up exactly to the change. Details are in App. A.9 and A.10. We note that this only works on algorithmic tasks with known symmetries, and have added a discussion on estimating these components for the other setups (see the added “Causal check on the GG retained by implicit federation bias” paragraph in Section 5, and App. A.10).

Figure 9

Panel (a) was a centralized learning rate sweep, with no FedAvg baseline so it reported absolute delay, while panel b compared federated algorithms so speedups were relative to FedAvg. We acknowledge the two cannot be read against each other, and have decided to remove panel a since it was providing little value, and add Figure 30, App. E.2 to compare the speedup of generalization relative to FedAvg with and without federation.

Comparison of gradient steps

All settings are measured in gradient work in units of the training set, so a full-participation round of E local epochs counts as E centralized steps and partial participation counts only the samples actually processed. Runs are compared at equal total work against a centralized baseline with the same model, data, optimizer and seed (see App. A.7).

If this has addressed your concerns, particularly regarding the novelty and soundness of our findings, we’d be grateful if you could revisit your score. We also welcome additional questions and discussion as needed.

— The Authors

Dear Authors,

Thank you for your detailed response and the additional experiments. I appreciate your efforts to address my concerns, especially the clarifications regarding the effects of local data size and the functional decomposition.

While these additions help clarify several points, I still have concerns about the novelty and significance of the main contributions. In particular, as also noted by Reviewer WMeJ, I am not fully convinced that the proposed implicit federation bias offers sufficiently new insights beyond the expected effects of aggregation in federated learning. I also think much of the analysis focuses on generalization timing rather than providing a deeper understanding of emergent generalization itself.

Overall, while I appreciate the improvements, my main concerns remain, so I will keep my original rating.

Dear reviewer sy84,

Thank you for your reply.

Regarding your concerns about the significance of our contributions, our findings give a different perspective to emergent generalisation in federation settings, not concluding it as an effect due to the size of client drift but something else implicit to the federated regime itself. This is in contrast to many other works which often focus on client drift. Regarding your remarks about insights beyond expected effects of aggregation, first, to our knowledge no such work has investigated this area and showed such results, and second, our work helps to explain the distinct joint training-test accuracy rises, whilst also giving a unified framework that holds across differing levels of federation.

We appreciate your remarks about focusing on generalisation timing rather than other factors, and so previously added additional analysis in Section 5 and the appendices (see the parts linking high federation and keeping G component only). The main focus was kept on generalisation timing since weight norm and other analysis showed highly inconsistent behaviour despite similar accuracy curves and grokking behaviour (Section 5 and Appendices). We cover this in Section 5 and discuss how this implies a different mechanism to other emergent generalisation findings.

If this has addressed your concerns, we’d be grateful if you could revisit your score. We again welcome additional questions, discussion and suggestions.

— The Authors

Reviewer_eSv5OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how federated training affects emergent generalization in grokking tasks. Experiments on modular arithmetic, permutation composition, and MNIST examine the effects of client count, local updates, participation, data heterogeneity, and optimization algorithms. The results show that federation can change both the timing and trajectory of generalization. A functional analysis of modular addition suggests that aggregation suppresses client-specific changes while retaining globally consistent structure, and server-step controls show that update magnitude also affects generalization timing.

Critical strengths
  1. The paper studies generalization throughout federated training rather than only reporting final accuracy. The observation that training and test accuracy can improve together after a prolonged training plateau offers an interesting distinction from classical grokking.

  2. The authors evaluate client count, local updates, participation, data partitions, and federated optimization algorithms across five setups. The functional decomposition and server-gain experiments provide evidence that both aggregation filtering and update magnitude influence generalization dynamics.

Critical weaknesses
  1. The paper compares several federated optimization algorithms, but does not systematically compare optimizers within the same centralized setting. Figure 9a studies centralized learning rates, which is useful, but leaves unclear which differences are specific to federation and which reflect more general optimizer effects on grokking. The interpretation of the optimizer results should distinguish these possibilities more clearly.

  2. The relationship between aggregation, weight norms, and grokking remains insufficiently explored. Comparing centralized and federated weight-norm trajectories, together with client norms before aggregation and the global norm afterward, would help connect the proposed filtering mechanism to existing norm-based explanations. Setup A uses no weight decay, but the sensitivity of the AdamW-based results to weight decay is not examined.

  3. The experiments focus on controlled algorithmic tasks and small-data MNIST. It remains unclear whether the observed phenomena extend to larger models and more realistic federated tasks. In particular, the detailed functional explanation relies on modular-addition structure, so its applicability beyond this setting should be stated more carefully.

  4. The practical implications for federated training remain underdeveloped. Although the paper explains that update scale can affect generalization timing, it provides limited guidance on how the findings could inform training duration, local-update schedules, or optimizer selection. It would also be useful to distinguish earlier generalization in normalized training time from actual communication or computational savings.

Questions for authors

See weaknesses.

2 comments

Dear Reviewer eSv5,

Thank you for your review and for your remarks on the observed delayed joint training and test increases being an interesting distinction from classical grokking. Below we address your specific concerns and raised questions.

Centralized versus federated optimizer effects

We have conducted additional ablations on the effect of optimizers in centralized settings (see Appendix E.2, Figure 30), to address concerns about the effect of optimizers rather than federation.

The results strengthen our interpretation of two complementary ways to accelerate emergent generalization (see Section 5 “Server step gain biases learning and generalization” paragraph). The first is the effective scale at which updates are applied: the speedups from a larger server step, FedAvgM and FedAdam carry over in full from the single-client case to ten clients, so federation does not diminish them. The second is the structure of the updates that survive aggregation, which is the subject of the rest of our work on implicit federation bias. Given that all but the FedYogi data point on the plot sit above the y=x axis, we can infer from this that the optimizer differences alone do not fully explain the effects of FL on convergence speeds.

Norms, weight decay and additional analysis

We thank the reviewer for their feedback and have now added additional figures and analysis on weight norms, as well as connecting our findings directly to existing grokking explanations (norm-based as well as other mechanisms) (see Section 5, paragraph “Relation to other grokking explanations”, and Appendices D.1 and D.2).

Weight decay was not investigated initially since the setups were taken widely from a variety of grokking works to give a broad set of settings to investigate from, and so are not the focus. We have swept the decay on all four AdamW setups. In centralized training, as expected stronger decay brings generalization forward and lowers the norm at which it occurs, and without decay Setups B–D do not generalize at these training fractions. However, the cost of federation is largely insensitive to the decay. For example on Setup C, FedAvg with 10 clients takes 1.2× and 1.3× the centralized time at decay 0.1 and 1.0, and with 20 clients 2.1× at both. Federated grokking also does not require decay: at a larger training fraction, Setup C groks under FedAvg with no decay, exhibiting the similar behaviour as the number of clients increases.

Usage and practical implications

Thank you for this suggestion. We have added a causal oracle intervention (Section 5, Figure 10, Appendix A.10). Amplifying, or keeping only, the shared component GG retained by aggregation reaches generalization 1.9–3.1× sooner in Setup A. This confirms that GG points towards the generalizing solution and through it we can alter the federated optimizer by boosting the learning signal on shared generalizable (or conversely memorised) components to change the learning dynamics such as to reproduce the shown joint training-test increases, to speed up convergence and produce grokking-like behaviour whilst keeping weight norms low. However, because GG relies on the task’s symmetry, this is currently diagnostic rather than for practical use; we discuss initial work and directions on GG estimation to then use in optimizers.

We have also added some communication and parallel timing savings analysis as requested. (Appendix E.5, Figure 36).

Larger models and tasks

This paper investigated algorithmic and small datasets and models as they provide more tractable settings to investigate behaviour and this is common in the field and surrounding literature [1–5].

Regarding the extension to larger models and more realistic tasks, we acknowledge this in the limitations section (Appendix F). Extending our work to larger models and more realistic federated tasks is future work, however there is evidence that it is feasible; concurrent work in centralized learning [6] identifies what a model is learning from its internal representations and uses it to steer training on LMs of 7B-32B scale, which suggests interventions of this type are feasible there.

We would like to thank the reviewer since their feedback has helped produce additional interesting findings.

If this has addressed your concerns, we’d be grateful if you could revisit your score. We also welcome additional questions as needed.

— The Authors

References:

[1] Liu et al. Towards Understanding Grokking: An Effective Theory of Representation Learning. NeurIPS, 2022.

[2] Nanda et al. Progress Measures for Grokking via Mechanistic Interpretability. ICLR, 2023.

[3] Liu et al. Omnigrok: Grokking Beyond Algorithmic Data. ICLR, 2023.

[4] Stander et al. Grokking Group Multiplication with Cosets. ICML, 2024.

[5] Kumar et al. Grokking as the Transition from Lazy to Rich Training Dynamics. ICLR, 2024.

[6] Fel et al. Shaping and Monitoring Models Through Their Internal Representations. arXiv 2026.

Reviewer_eSv5OpenReview ↗

Response to rebuttal

I thank the authors for their response. I am inclined to maintain my positive assessment of the paper and encourage the authors to extend their findings to larger models and a broader range of tasks in future work.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

Dear Reviewers, ACs, and PCs,

We thank each of the reviewers for their time and constructive feedback given. To address each of the points raised, we have responded to each reviewer directly, noting changes and items addressed. We note that the majority of changes occur in the appendix and are extensions to previous work. This shared comment is written to record and detail each of the changes and revisions made to the paper, as well as the reasons for changes. We hope this provides greater transparency and assistance.

– The Authors

Revised Version #1

  1. Presentation and clarity. We altered the abstract and introduction to better communicate the paper's findings, novelty and contributions, alongside clarifications throughout the manuscript. (sy84: W1, W3)

  2. Discussion on existing grokking explanations. Expanded discussion connecting our findings to norm-based, circuit-based and other explanations of grokking (Section 5). (sy84: W1; WMeJ: Q1; eSv5: W2)

  3. Functional decomposition details. Expanded methodological details on extracting functional components and discussed potential extensions beyond tasks with known symmetries (Appendix A.10). (sy84: Q2; eSv5: W3)

  4. Data-size controls. Added experiments distinguishing federation effects from reductions in data available per client (Appendix D.3, Figures 25–26). (sy84: W2, Q1)

  5. Causal interventions. Added experiments using the functional analysis to alter generalization speed and learning dynamics through selective amplification or retention of GG (Section 5, Figure 10; Appendix A.10). (WMeJ: Q2; sy84: W1; eSv5: W4)

  6. Weight-norm details. Added global weight-norm trajectories across federated configurations and optimizers (Appendix D.1, Figures 13–18). Also added trajectories of client weight-norm variation and movement during local optimization and aggregation (Appendix D.2, Figures 19–24). (eSv5: W2)

  7. Centralized optimizer comparisons. Added controls distinguishing general optimization effects from federation-specific effects (Appendix E.2, Figure 30). (eSv5: W1)

  8. Analysis on communication and computational savings Added analysis of communication requirements, computational work and idealized parallel training time (Appendix E.5, Figure 36). (eSv5: W4)

  9. Figure 9 revision. Revised Figure 9 to improve clarity, added Figure 30 to compare federated and non federated algorithm convergence speed. (sy84:Q3, eSv5:W1)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.