acceptodds
Under review as a conference paper at ICLR 2027

Whom to Trust? Adaptive Collaboration in Personalized Federated Learning

Abstract

Data heterogeneity poses a fundamental challenge in federated learning (FL), particularly when clients differ in both their data distributions and the reliability of their predictions. We observe that many personalized FL (PFL) methods fail to outperform the essential local and centralized baselines, revealing a meaningful personalization regime in which a single global model is insufficient, yet cross-client collaboration remains useful. From this diagnosis, we identify two requirements for effective personalization: adaptivity in how strongly each client relies on collaboration and fine-grained selectivity in whose predictions should be trusted. We instantiate these principles in FEDMOSAIC, a personalized federated co-training method that dynamically balances local and collaborative objectives while weighting client contributions at the example level. Across label-skew, feature-shift, and hybrid settings, FEDMOSAIC outperforms strong FL and PFL baselines and frequently surpasses both local and centralized training. We further establish convergence under standard smoothness, bounded-variance, and pseudo-label drift assumptions, and characterize a client-level differentially private variant of the protocol. These results clarify when federated personalization is useful and how adaptive, selective collaboration can realize its benefits.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Personalized Federated Learning, Co-training, Federated semi-supervised Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_sS9nOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents an adaptive personalized FL method to address extreme non-IID environments.

Critical strengths
  • This paper has a good presentation, and the method is well defined.

  • The author has provide a decent appendix list.

Critical weaknesses

Weaknesses:

  • In the Introduction (Lines 43–44), the authors state that “Tab. 1 shows that in heterogeneous regimes both FL and even centralized training perform worse than local training.” I do not find this claim sufficiently reliable. The results in Table 1 are obtained under an extremely strong label distribution shift, which may partially explain why FedAvg performs substantially worse than even the centralized setup. Under such an extreme label-support mismatch, strong local performance and severe degradation of a single global model are largely expected. Therefore, I do not think Table 1 provides convincing evidence for the broader claim that FL or centralized training generally underperform local training in heterogeneous regimes. The severity of the non-IID setting, particularly the label distribution shift in this case, appears to be a key factor driving these performance gaps.

  • Personalized FL is commonly used to address client heterogeneity. However, another line of work, clustered FL, also targets the same problem. I am not suggesting that more baselines are necessarily required, but I think it would be important to compare the proposed method with recent clustered FL approaches, ideally from 2025–2026. Ultimately, from a practitioner's perspective, the question is which method should be chosen to address this type of heterogeneity. The empirical evaluation should therefore provide stronger evidence that the proposed pFL method performs better than clustered FL approaches for the problems considered in this work.

  • Related to the first comment, the P(Y)P(Y) partition used in this work (two classes per client) already represents an extremely strong non-IID setting. As a result, a substantial portion of the headline empirical evidence, particularly Table 1 and the pathological label-skew experiments, may have limited external validity.

  • Considering the comments above, I find that the current empirical evaluation is not sufficiently strong to support the paper's broad conclusions. In my view, the paper would require a substantially stronger and more systematic experimental setup to adequately support these claims, along with a stronger baseline selection.

Questions for authors

Please see Critical Weaknesses.

4 comments

We thank the reviewer for raising the question of external validity under severe heterogeneity.

Severity of the non-IID setting

We agree that the two-class-per-client partition represents extreme label heterogeneity. We also agree that the statement motivated by Table 1 is too broad. The table demonstrates the existence of a regime where centralized and conventional federated training can underperform local training. It does not imply that this behavior is typical of heterogeneous FL. We will clarify the wording.

However, the review appears to interpret this pathological partition as the main basis for our conclusions. This is not the case.

Section 4 evaluates three different heterogeneity scenarios and explicitly states that “each scenario stresses a different aspect of adaptive collaboration” (lines 318-323):

  • Pathological and practical Dirichlet label skew, where the latter allows overlapping client label distributions.
  • Pure feature shift on Office-10 and DomainNet, without imposing the two-class-per-client partition.
  • Hybrid label and feature shift.

The pure feature-shift experiments are particularly relevant to the reviewer's concern, since they do not rely on the extreme label-support mismatch criticized in the review. Table 3 shows that FEDMOSAIC improves over both local and centralized training in these settings. Averaging the reported per-domain accuracies gives:

DatasetCentralizedLocalFEDMOSAIC
DomainNet61.5561.55%65.0165.01%70.1570.15%
Office-1072.4872.48%66.8966.89%76.1376.13%

Thus, the improvement is not limited to the pathological label-skew setting. In fact, the gains are considerably more meaningful here than in the nearly saturated pathological Fashion-MNIST and CIFAR-10 experiments.

The practical Dirichlet experiments provide further evidence. On CIFAR-10, FEDMOSAIC obtains 86.1586.15% compared with 83.9183.91% for local training. On Tiny-ImageNet, it obtains 41.9041.90% compared with 36.7536.75%. Centralized training remains slightly better in these two cases.

This qualification is already made in the Discussion. We explicitly state that the gain over local training under pathological label skew is very small, question whether such settings are particularly informative for collaborative gain, and note that under the milder practical non-IID setting “centralized training performs best, as expected” (lines 460-476).

Our claim is therefore not that centralized training generally fails under heterogeneity, nor that FEDMOSAIC always outperforms centralized learning. We study the regime described in the paper where a single shared model is insufficient but collaboration can still improve personalized models.

Clustered FL

We agree that clustered FL is an interesting and relevant direction. We see it, however, as largely orthogonal to the problem studied here.

Client clustering determines which groups of clients should collaborate. Our expertise mechanism asks whose prediction should matter for this particular example within a collaborating set. These are different decisions. Two clients can be useful collaborators for one part of the input space and poor collaborators for another, even if they belong to the same cluster.

The two approaches are also naturally compatible. Clustering could restrict which clients participate in a consensus, while expertise weighting determines their relative influence for each example within that cluster. Alternatively, different clusters could construct separate consensuses. This would add another layer of personalization to the collaboration structure rather than replace the mechanism studied here.

Our evaluation already includes PerFed-CKT, which uses clustered knowledge transfer. In the hybrid experiments (Table 4), FEDMOSAIC-W achieves 87.4487.44% on DomainNet and 89.0689.06% on Office-10, compared with 79.2479.24% and 82.4982.49% for PerFed-CKT.

We agree that a broader study combining recent clustering approaches with example-level expertise would be interesting future work. It introduces an additional question, namely which clients should collaborate at all, whereas the present paper focuses on how collaboration should behave once a set of clients collaborates: when to trust the collaborative signal and whose predictions to trust at the example level.

In summary, pathological label skew alone would indeed provide limited evidence for general applicability. The submitted evaluation, however, also includes practical label skew, pure feature shift, and hybrid heterogeneity. The pure feature-shift results in particular demonstrate substantial collaborative gains without the extreme label-support mismatch on which the review focuses.

Dear Reviewer sS9n,

As the rebuttal deadline approaches, we would like to briefly follow up. We have clarified the scope of our claims and highlighted that the original submission already includes practical Dirichlet label skew, pure feature shift, and hybrid settings, demonstrating gains beyond extreme non-IID scenarios.

We also clarified our distinction from clustered FL and highlighted the existing PerFed-CKT ( clustering approach ) comparison.

We hope these explanations address your concerns and would appreciate your reconsideration of the evaluation.

I thank the author for the rebuttal materials, and they have mostly addressed my concerns. I will update my score.

We sincerely appreciate your constructive engagement throughout the discussion and thank you for reconsidering and increasing your evaluation score. We are pleased that our clarifications addressed your main concerns.

Reviewer_96XqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates how personalized federated learning can benefit from collaboration under heterogeneous client distributions without degrading strong local models. FedMosaic extends federated co-training with two mechanisms: example-level expertise-weighted aggregation of hard predictions on a shared unlabeled dataset, and a client-specific coefficient controlling reliance on the resulting pseudo-labels. Each client optimizes:

Li=Llocal,i+λiLpublic,i\displaystyle \mathcal{L}_{i}=\mathcal{L}_{\mathrm{local},i}+\lambda_{i}\mathcal{L}_{\mathrm{public},i}

λi=exp⁡(1−Lpublic,iLlocal,i)\displaystyle \lambda_{i}=\exp\left(1-\frac{\mathcal{L}_{\mathrm{public},i}}{\mathcal{L}_{\mathrm{local},i}}\right)

Expertise is estimated using class frequency or predictive uncertainty. The server broadcasts the same consensus labels to all clients, while their models and reliance coefficients remain personalized. The evaluation covers label skew, feature shift, and hybrid heterogeneity, with additional component ablations and robustness controls. The paper also provides an optimization analysis under bounded objective drift and a conditional client-level differential privacy analysis.

Critical strengths
  1. FedMosaic integrates example-level expertise weighting with client-specific adaptive reliance, separating consensus construction from how strongly each client uses that consensus. This is a useful algorithmic extension of hard-label federated co-training, with ablations supporting the complementarity of the two mechanisms under hybrid heterogeneity.

  2. The design is lightweight, communicating only hard predictions and one expertise score per public example rather than model parameters or full probability vectors. It retains locally personalized models while offering substantial communication savings when the public dataset is small relative to the model size.

Critical weaknesses
  1. The adaptive rule in Equation 3 can reinforce harmful collaboration, so the general fail-safe claim is not established. Its advantage over a well-tuned fixed coefficient also remains unclear. The shared consensus does not implement recipient-specific peer selection, and the methodological distinction from FedCT and FedPD needs sharper positioning.

  2. Tables 3, 5, and 9 contain inconsistent configuration labels and results, limiting reproducibility and component attribution. Proposition 1 also uses full-dataset variance-reduction factors without sufficient justification for the stated mini-batch estimator. These reporting and analytical issues require clarification or correction.

References:

[1] Abourayya et al. "Little Is Enough: Boosting Privacy by Sharing Only Hard Labels in Federated Semi-Supervised Learning." AAAI, 2025. Relevant location: Appendix D.6. https://arxiv.org/html/2310.05696v4

[2] Yang et al. "FedPD: personalized federated learning based on partial distillation." Frontiers of Computer Science, 2026. Relevant mechanisms: Partial Knowledge Transfer and Partial Knowledge Ensemble. https://doi.org/10.1007/s11704-025-40840-4

Questions for authors
  1. Are there safeguards or explicit conditions that exclude the feedback failure described above? In one representative hybrid setting, does the adaptive coefficient outperform a tuned fixed coefficient under the same consensus and training budget? Joint loss/coefficient trajectories and per-client gains would help establish its independent value.

  2. Does the FedCT baseline use ordinary majority voting or the qualified-majority variant? Relative to FedPD, what distinct capability or efficiency advantage does the shared hard-label consensus provide? A mechanism-level explanation is sufficient; an exhaustive baseline suite is not required.

  3. Please reconcile Tables 3, 5, and 9 and specify the active components and exact expertise-score formulas. What sampling assumptions justify the full-dataset variance factors in Proposition 1 for the implemented mini-batch estimator?

6 comments

We thank the reviewer for the detailed technical comments. We agree that several aspects require more precise statements.

Adaptive weighting and the fixed-λ\lambda comparison

Eq. 3 has an explicit suppression property. Writing $

rit=Lglobal,itLlocal,it,r_i^t=\frac{\mathcal L_{\mathrm{global},i}^t} {\mathcal L_{\mathrm{local},i}^t},

$ Eq. 3 gives $

λit=exp⁡(1−rit).\lambda_i^t=\exp(1-r_i^t).

$ Hence, for any (ε∈(0,1)\varepsilon\in(0,1)), $

rit≥1+log⁡(1/ε)⟹λit≤ε.r_i^t\geq 1+\log(1/\varepsilon) \quad\Longrightarrow\quad \lambda_i^t\leq\varepsilon.

$ Thus, whenever the global loss is sufficiently larger than the local loss, the collaborative objective is exponentially suppressed. This is exactly the behavior examined in Appendix D for the client with incompatible label semantics, where the global loss is substantially higher than the local loss. The corresponding performance effect is visible in Fig. 6: for the incompatible client, fixed expertise-weighted collaboration yields 164/300164/300 errors, whereas adding adaptive collaboration reduces this to 15/30015/300, matching local training.

This is not an unconditional guarantee against every possible form of harmful collaboration. Incorrect pseudo-labels may in principle also attain low loss, in which case the loss ratio alone cannot identify the failure. Our use of “fail-safe” in the public-data experiment was intended to describe the observed behavior in that setting, not to state such a universal guarantee. We have updated the manuscript accordingly and changed the fail-safe claim to “ mitigating negative transfer in the evaluated settings”.

Regarding the fixed-coefficient comparison, it is important to distinguish a single fixed coefficient shared across clients from a client-specific fixed coefficient. A single fixed coefficient cannot represent the controlled setting in Fig. 1, where the regular clients benefit from collaboration while the incompatible client should largely suppress it.

A separately tuned fixed (λi\lambda_i) for every client could represent this difference. However, this would itself be an implementation of the client-specific reliance principle we advocate. It would also require a representative validation set and a separate hyperparameter search for each client. Eq. 3 provides an online realization using quantities already available during training. Our claim is therefore not that (λi\lambda_i) must change over time, nor that Eq. 3 must outperform every oracle-tuned fixed (λi\lambda_i), but that collaboration strength should adapt to the recipient.

We also agree that the consensus itself is shared rather than recipient-specific. The precise distinction is:

  • example-specific contributor weighting: whose prediction should shape the pseudo-label for this example?
  • recipient-specific reliance: how strongly should client i use the resulting pseudo-label? This is the same distinction made in Sec. 3, lines 163-173.

Relation to FedCT and FedPD

The FedCT baseline uses standard majority voting. We have updated the manuscript and state this explicitly.

Relative to FedCT, FEDMOSAIC modifies both consensus construction and how the resulting consensus is used. FedCT treats contributors uniformly. FEDMOSAIC instead weights contributor expertise separately for each public example. Each recipient then independently controls the influence of the resulting consensus on its local objective.

FedPD also addresses selective knowledge transfer, but at a different level. Its Partial Knowledge Transfer mechanism selects knowledge according to its relevance to the recipient, while Partial Knowledge Ensemble maintains personalized server-side models for individual clients.

FEDMOSAIC instead asks a contributor-side question before the recipient uses the collaborative signal: which clients are reliable for this particular public example? Their contributions are weighted when constructing a shared hard-label consensus. Recipient-side adaptation is then handled separately through λi.\lambda_i.

The distinction is therefore not simply that one method is selective and the other is not. FedPD performs recipient-oriented selective distillation. FEDMOSAIC explicitly separates example-specific contributor expertise from recipient-specific reliance within hard-label federated co-training. It also does not require personalized server models and communicates hard predictions plus one expertise value per public example rather than full prediction distributions. We have revised the paper and added a detailed discussion comparing FedCT and FedPD in the Extended Related Work section of Appendix C.

Tables 3, 5, 9 and Proposition 1

We thank the reviewer for identifying the configuration inconsistency. The underlying results in Tables 3 and 5 are consistent with the experimental design described in Sec. 4 (lines 318–323). Each main experiment uses the mechanism most directly motivated by the corresponding heterogeneity setting: class-frequency expertise weighting for label skew (Table 2), adaptive collaboration through dynamic (λ\lambda) for feature shift (Table 3), and both mechanisms for hybrid heterogeneity (Table 4). In particular, the DomainNet results in Table 3 average to 70.1570.15%, exactly matching the (λ\lambda)-only configuration in Table 5. Enabling both mechanisms under feature shift increases this to 70.8070.80% in Table 5. Likewise, the 87.4487.44% hybrid DomainNet result in Table 4 matches the full configuration in Table 5.

The inconsistency is in the description and configuration labels surrounding Table 9, which incorrectly characterize the Table 3 results as using the class-frequency expertise mechanism. We thank the reviewer for catching this labeling error. We have corrected Table 9 and the preceding text and explicitly state the active components and expertise-score definition for every reported configuration.Furthermore, we now explicitly define both expertise-score mechanisms in Appendix A (page 15, lines 782–809).

Regarding Proposition 1, the reviewer is correct that the variance terms should use the mini-batch sizes rather than the full dataset cardinalities. Under the assumed mini-batch sampling, the corresponding variance terms are (σ2/blocal\sigma^2/b_{\mathrm{local}}) and (σ~2/bglobal\tilde{\sigma}^2/b_{\mathrm{global}}), respectively. We have updated the manuscript and corrected these denominators in Proposition 1 and Appendix A.

Dear Reviewer 96Xq,

As the rebuttal deadline approaches, we would like to briefly follow up. We have clarified the adaptive weighting mechanism and its limitations, strengthened the comparison with FedCT and FedPD, and corrected lable of Tables 9, as well as the variance terms in Proposition 1.

We hope these revisions and explanations address your concerns and would appreciate your reconsideration of the evaluation.

Reviewer_79L6OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary
  • Setting: In (personalized) federated learning, there are many clients (local learners). Each client has its own local data and may exchange information with other clients. A local baseline means each client learns with no information exchange. A centralized baseline means a centralized learner learns with all data from all clients. In some scenarios, federated learning may not outperform both baselines. This paper proposes methods to improve federated learning performance, aiming to outperform both baselines.
  • Methods: This paper lets clients exchange information explicitly through the pseudo-label assignment of a public unlabeled dataset. Particularly, the pseudo-label of each public sample is weighted majority voted by the predictions of all clients, where the weight reflects the reliability of clients' predictions. Each client minimizes a weighted combination of its local loss and a public loss based on the pseudo-labels, where the weight is dynamically adjusted according to the local loss and the public loss, reflecting the alignment between the global consensus and the local knowledge.
  • Results: This paper conducts experiments and shows that proposed methods outperform both baselines in most of these experiments. This paper also gives convergence analysis and differential privacy analysis under some assumptions.
Critical strengths
  • Significance: The problem this paper addresses is fundamental in federated learning. Different clients may have complementary information that can improve the performance of some other clients. Inappropriate information exchange may hurt the performance of some clients due to unreliable information or misalignment between global consensus and local knowledge.
Critical weaknesses
  • Limited novelty: This paper heuristically extends FedCT with special treatment of sample-wise weighting and round-wise weighting, following intuitive ideas. It is less clear what new insights these heuristic extensions provide into the "data heterogeneity" setting beyond their intuitive motivations. While the general motivation of this paper is reasonable, a stronger justification of the proposed methods would be helpful to clarify the methodological contribution of this paper.
Questions for authors

None.

4 comments

We thank the reviewer for recognizing the importance of the problem and for finding no substantive soundness concerns.

The main question is what new insight the proposed “sample-wise weighting and round-wise weighting” provide beyond their intuitive motivation. At the level of the individual mechanisms, the reviewer is right that both weighting ideas are intuitive. The new insight is that they address two different failure modes of collaboration, and that their relative importance changes with the structure of heterogeneity.

The paper therefore does not rely on the novelty of weighting itself. Instead, it asks what is missing when personalized FL fails to obtain genuine collaborative gain over strong local models. Our experiments point to two distinct problems:

  • First, the reliability of contributed knowledge depends on the example. Under label skew, a client may be a useful expert for classes represented in its local data but unreliable for others. Uniform aggregation ignores this distinction. FEDMOSAIC therefore uses example-level expertise weighting when constructing the consensus.

  • Second, the usefulness of the consensus depends on the recipient. Even when pseudo-labels are useful for most clients, they may conflict with another client's local task. Fig. 1 includes exactly such a client with incompatible label semantics. Improving the consensus alone cannot solve this problem. The client must also be able to reduce its reliance on the collaborative objective.

These are different failure modes. A better consensus does not necessarily imply better personalization for every client, and adapting collaboration strength does not make the underlying consensus more reliable.

This distinction is already stated in the paper as the two questions “when to trust the global signal?” and “whose predictions to trust?” (Sec. 3, lines 163-173), and in the Introduction as the need to decide both how much to trust the collaborative signal and whose predictions to trust for each example (lines 83-88).

It also explains the ablation results in Appendix D, Table 5:

  • Under label skew, expertise weighting is the more important mechanism.
  • Under feature shift, adaptive reliance contributes more.
  • Under hybrid heterogeneity, both mechanisms contribute. On DomainNet, expertise weighting alone reaches 84.8084.80%, adaptive reliance alone 85.6085.60%, and the combined method 87.4487.44%.

The paper therefore does more than show that two intuitive weighting rules improve FedCT. It identifies two different sources of harmful collaboration, constructs experimental settings in which their roles differ, and shows that the importance of the two mechanisms changes with the underlying heterogeneity. FEDMOSAIC is one realization of this separation. We do not claim that the particular weighting functions are uniquely optimal or that these are the only possible implementations.

Dear Reviewer 79L6,

As the rebuttal deadline approaches, we would like to briefly follow up. Our response clarifies how FEDMOSAIC addresses two distinct failure modes of collaboration, with Table 5 demonstrating their different roles across heterogeneity settings.

We hope this addresses your novelty concern and would appreciate your reconsideration of the evaluation.

Thank you for your detailed reply. Now I have a better understanding of the two failure modes: one concerning the contributor (on each sample) and the other concerning the recipient (at each round). This paper proposes separate treatments for these modes, and the empirical results show that each treatment can help mitigate corresponding issues in PFL, while combining them leads to further improvements.

Dear Reviewer 79L6,

Thank you for your thoughtful follow-up. We are glad that our response clarified the two distinct failure modes and how their complementary treatment contributes to personalized federated learning.

Beyond the empirical evidence supporting this distinction across 5 benchmark datasets, 3 heterogeneous scenarios, and comparisons with 12 baselines, we also provide theoretical convergence guarantees and a conditional client-level differential privacy analysis, further supporting the proposed framework.

We sincerely appreciate your constructive engagement throughout the discussion and thank you for reconsidering and increasing your evaluation score. We are pleased that our clarifications addressed your main concerns, and we are grateful for your positive reassessment.

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

This paper studies personalized federated learning (pFL) under heterogeneous client distributions. The main motivation is based on the observation that existing pFL approaches may fail to outperform either local or centralized training, suggesting the need to control the degree of collaboration among clients to achieve the best accuracy. Specifically, the authors propose a mechanism to decide both when a client should collaborate and which other clients should be trusted. The paper proposes FedMosaic, a personalized federated co-training method built on FedCT. In each round, clients make predictions on a shared, unlabeled public dataset. The server uses each client’s prediction confidence vectors to generate global consensus pseudo-labels for all samples in the public dataset, which are distributed back to the clients. Each client trains its personalized model using its local dataset and the additional shared pseudo-labeled dataset, with dynamic collaboration weights. The method is evaluated under different heterogeneity settings against several FL and pFL methods. The results show significant accuracy improvements compared with competing methods and outperform the baselines, including centralized and local training, in some cases.

Critical strengths

S1. Strong empirical performance under severe heterogeneity with broad experimental evaluation and setting. S2. The motivation points out by the paper is interesting and important in pFL research area.

Critical weaknesses

W1. Limited novelty. The proposed method is based primarily on FedCT. The key idea of allowing clients to exchange predictions on a shared, unlabeled public dataset was already introduced in the FedCT paper. FedMosaic appears to extend FedCT mainly through adaptive local-global loss weighting and confidence-weighted pseudo-labeling. The authors should clarify how FedMosaic differs from simply combining FedCT with established adaptive personalization and confidence-weighting techniques.

W2. Heavy dependence on public data. The method depends heavily on the availability and quality of the public dataset. Figure 3 shows that performance is sensitive to public-data quality. In practice, collecting a representative public dataset with characteristics similar to those of the clients may be difficult. Although the authors argue that public datasets are available in many application domains, the effectiveness of FedMosaic when using an independently collected public dataset remains unclear. In the experiments, the public dataset is created by extracting a small subset of the dataset, e.g., CIFAR-10 used for evaluation. A more realistic experiment would use a separate dataset as the public data, such as a subset of ImageNet for CIFAR-10 or Tiny ImageNet experiments. Table 3 may partially address this concern, but the authors should explain the setting and its implications more clearly.

W3. Limited evaluation of client participation and scale. The authors appear to assume full client participation in every communication round. They should discuss and evaluate FedMosaic under participation rates below 100%. Moreover, the experiments involve relatively few clients, limiting the evidence for scalability. The authors should also clearly report the Dirichlet distribution parameter used to generate non-IID data.

Questions for authors

See above

5 comments

We thank the reviewer for the positive assessment of the motivation and empirical evaluation. We address the three concerns below.

W1: Novelty relative to FedCT

We agree that prediction exchange on a shared unlabeled dataset is not new. FEDMOSAIC explicitly builds on FedCT. Our contribution is to identify and address two distinct problems that arise when using this information for personalized learning:

  • Whose predictions to trust? Clients may be reliable on some examples but not others. We therefore weight their contributions individually for each public example, rather than treating all predictions equally.

  • When to trust the consensus? Even a reliable consensus may be incompatible with a particular client's local task. We therefore let each client independently adjust how strongly it uses the collaborative objective.

These are the two questions stated explicitly in Sec. 3 (lines 163-173). They also correspond to the two failure modes illustrated in Fig. 1: uniform prediction sharing can be corrupted by non-experts, while a single collaboration rule can be harmful for an incompatible client (lines 83-88).

The two mechanisms therefore address different problems. Improving the consensus does not imply that every client should use it equally, while adjusting collaboration strength cannot compensate for systematically unreliable pseudo-labels.

The experiments support this distinction. Sec. 4 explicitly states that each heterogeneity scenario stresses a different aspect of adaptive collaboration (lines 318-323). Appendix D, Table 5 then shows the corresponding mechanism dependence. Under label skew, expertise weighting is particularly important. Under feature shift, adaptive reliance contributes more. Under hybrid DomainNet heterogeneity, expertise weighting alone achieves 84.8084.80%, adaptive reliance alone 85.6085.60%, and their combination 87.4487.44%.

We do not claim that either weighting mechanism is individually new. The contribution is their separation into two complementary collaboration decisions, together with an empirical study of when each becomes important. This is also how the paper positions the contribution in lines 148-153: FEDMOSAIC is one concrete realization of personalization as selective collaboration.

The paper additionally characterizes this realization through a convergence analysis for dynamically weighted objectives with evolving pseudo-labels (Proposition 1, Appendix A) and a conditional client-level privacy analysis for the communicated labels and expertise scores (Proposition 2, Appendix B).

W2: Dependence on public unlabeled data

The shared unlabeled dataset is not an additional assumption introduced specifically by FEDMOSAIC. It is a standard setting in federated semi-supervised learning. Distributed Distillation (Bistritz et al., 2020) already uses predictions on unlabeled public data for federated collaboration, and FedCT (Abourayya et al., 2025), on which FEDMOSAIC directly builds, uses a shared unlabeled dataset as its communication substrate. The paper states this motivation explicitly in the Introduction (lines 96-102).

Thus, requiring U is part of the federated semi-supervised/co-training setting we study rather than a new assumption introduced to obtain FEDMOSAIC's gains. In particular, FedCT receives the same public-data substrate as FEDMOSAIC. The substantial performance difference between the two methods therefore cannot be explained simply by access to U.

We nevertheless agree that the quality and availability of UU are important practical questions, which is why we study them directly. Appendix D varies |U| from 3,0003,000 to 250250 examples and evaluates increasingly skewed public distributions down to Dirichlet α=0.1\alpha=0.1, including cases where several classes are absent. The appendix also explicitly discusses public-data availability as an assumption of federated semi-supervised learning (lines 1367-1375), and the Discussion lists unavailable public data as a limitation (lines 478-482).

These experiments demonstrate robustness to public-data scarcity and substantial class-distribution mismatch. They do not establish robustness to arbitrary source shift from a completely independently collected dataset, as the reviewer correctly points out. We will clarify this distinction.

W3: Partial participation and scalability

The requested experiments are already included in Appendix D. With 3030 clients, FEDMOSAIC achieves 93.1293.12%, 90.3290.32%, 89.2489.24%, and 88.4188.41% accuracy under dropout rates of 00%, 1010%, 2020%, and 3030%, respectively.

We additionally evaluate 100100 clients with only 2020% participation per communication round. FEDMOSAIC achieves 87.6487.64%, compared with 86.5186.51% for local training, 75.9175.91% for FedCT, and 69.2369.23% for FedAvg. The corresponding setup and controls are described in Appendix D, lines 1280-1295.

Thus, the submission already evaluates both partial participation and a substantially larger client population. We have updated the paper accordingly and also reported the Dirichlet concentration parameter used for the practical client partitions explicitly: α=0.1\alpha=0.1.

Dear Reviewer GVtf,

As the rebuttal deadline approaches, we would like to briefly follow up. We have clarified FEDMOSAIC's novelty over FedCT and addressed concerns about public-data dependence.

Importantly, the requested partial-participation and scalability experiments were already included in Appendix D of the original submission.

We hope our responses address your concerns and would appreciate your reconsideration of the evaluation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal Outcomes and Reviewer Discussion

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their constructive feedback and engagement during the rebuttal period.

We would like to briefly summarize the main outcomes of our discussion:

  • Novelty and methodological contribution: We clarified the two complementary failure modes addressed by FEDMOSAIC (example-level contributor reliability and client-specific reliance ) supported by extensive ablations across three heterogeneity scenarios.
  • Experimental evidence: We highlighted evaluations across five datasets, including public-data robustness, different heterogeneity settings, and partial-participation and scalability experiments that were already included in the original submission.
  • Technical clarifications: We strengthened comparisons with FedCT, FedPD, and clustered FL, corrected experimental configuration descriptions, and refined the convergence analysis and theoretical claims.

We are particularly encouraged by the reviewers' responses:

  • Reviewer 79L6 increased their score from 2 to 3, explicitly acknowledging the two distinct failure modes and the empirical evidence supporting their complementary treatment.
  • Reviewer sS9n increased their score from 1 to 3, stating that the rebuttal had mostly addressed their concerns.
  • Reviewer GVtf maintained their positive score of 3, explicitly confirming that most of their concerns had been fulfilled.
  • Reviewer 96Xq increased their score from 2 to 3 confirming that the rebuttal had largely addressed their technical concerns.

We believe the discussion has substantially clarified the paper's contributions, strengthened its technical presentation, and addressed the majority of concerns raised during review.

We sincerely appreciate the reviewers' time and thoughtful reassessments.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.