acceptodds
Withdrawn from ICLR 2027

Smooth Client–Instance Adapter Fields for Personalized Federated Learning under Compound Heterogeneity

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Personalized federated learning (PFL) commonly binds one model to each client or selects among finite experts, creating artificial boundaries between nearby distributions and reducing a multimodal client to one identity. We introduce FedSPLINE (Federated Smooth Personalization via Layerwise Implicit Neural Experts), a tensor-product cubic Bernstein field of low-rank adapters over learned client–instance coordinates. A distribution signature anchors each client, a bounded router refines instance queries, and sixteen nonnegative analytic weights blend shared controls while offsets and heads remain private. This explicit convex geometry gives partition-of-unity and Lipschitz properties, a cold-start decomposition, and supporting conditional shared-parameter analysis. Under source-identity-disjoint evaluation on three compound-shift benchmarks and twelve controlled baselines, FedSPLINE has the highest observed mean accuracy (83.11%) and label-free cold-start mean (63.23%). Its 0.14-point three-seed and 0.19-point five-seed HyperAdapter gaps are descriptively unresolved ( and ), while it remains Pareto-relevant through explicit geometry and slightly lower communication than the capacity-matched generator. On unseen interior coordinates, its 0.49-point advantage over that generator is nonsignificant (), and extrapolation is weaker. The evidence supports an inspectable geometry–utility trade-off, not a universally dominant PFL rule.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
personalized federated learning, compound heterogeneity, client-instance personalization, continuous parameter fields, low-rank adapters, spline interpolation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_eudxOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper focuses on the personalized federated learning and introduce federated smooth personalization via the layerwise implicit neural experts to mitigate the distribution shift. It introduce a lot of modules including a bounded rounter, sisteen nonnegative analytic weights, cold-start decomposition and conditional shared-parameter analysis.

Critical strengths
  1. Understanding the personalized federated learning from the geometric perspective is interesting.
  2. There is a comprehensive empirical evaluation for the paper.
Critical weaknesses
  1. Many words are too hard to understand. For example, only from the abstract, what is the bounded rounter, nonnegative analytic weights, supporting conditional shared-parameter analysis, source-identity disjoint evaluation etc. Similar for other parts of the paper, the readability of this paper is poor. Similar for the figures, it seems too complex to be understandable.
  2. The formulation of the personalized federated learning and the study self-defied compound heterogeneity is not clear. Is there any difference from the conventional PFL and the compound heterogeneity? Also it is ambiguous of what the P_i is defined here when you already have the D_i.
  3. Many words are very weird, for example “We do not claim cryptographic privacy for s_i xxx” in line 142. As a reader, one don’t even know what the cryptographic privacy means in this paper while the paper says it does not claim cryptographic privacy.
  4. In line 149, the paper mentions they have four design questions for PFL while this is not cleared before the section 4. In conventional PFL, there is no coordinate encoder (a convention FL just needs a learnable neuron network) and why it turns to be a design question in section 4.1 now? Similar for the other design questions.
  5. There are many modules introduced in this paper but there lack a motivation of how each module is brought in. Given the readability of the English, figure and math, it is very hard to understand the how each module works.
Questions for authors
  1. Could the authors please polish the readability of this paper to help improve the reader's overall experience as pointed by weakness 1-2. For the rebuttal, could you please provide the clairfication?
  2. Could the authors please explain why there are many weird statement and new jargon in the paper.
  3. Could the author please provide the motivation of those many modules introduced to the paper.
  4. Could the author please clarify their setting from conventional PFL.
Reviewer_wP6aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FedSPLINE, which learns a continuous field of low-rank adapters for personalized federated learning. Client distribution statistics and instance-dependent routers determine layerwise coordinates, which yield Bernstein weights for mixing shared adapter controls. Classifier heads and client offsets remain private. The paper provides geometric analysis and evaluates personalized accuracy, unseen-client cold start, interpolation, and robustness under compound label and feature heterogeneity.

Critical strengths
  1. The client–instance queries and convex Bernstein mixing provide an explicit way to construct sample-dependent adapters, with analyzable coordinate sensitivity. This connects client-level and within-client personalization in a coherent architecture.

  2. The method is evaluated across multiple datasets and settings, providing empirical support for its effectiveness.

Critical weaknesses
  1. The method combines several components, while related ideas, such as layerwise personalization and adaptive parameter mixing, have already been explored in PFL. The core contribution beyond these existing ideas is not sufficiently clear.

  2. It remains unclear why the adapter mixing weights induced by the proposed coordinate-based mechanism should be more effective than those produced by existing approaches.

  3. The distinctive Bernstein structure is not convincingly shown to improve prediction or robustness over HyperAdapter. The paper should explain more clearly what practical benefit the explicit geometry offers beyond an inspectable parameterization.

  4. The evaluation does not explicitly report the performance of each personalized model on a class-balanced test distribution while retaining its client-specific feature or domain conditions. This makes it difficult to distinguish improved feature adaptation from stronger specialization to local class priors.

  5. Classification performance is assessed primarily through accuracy, which does not adequately characterize minority-class performance under label imbalance. Additional metrics, such as macro-precision, macro-recall, macro-F1, and balanced accuracy, would provide complementary evidence. The paper should also clarify how client scores are aggregated and how absent classes are handled.

  6. Figure 7 shows anchors, queries, adapter-path behavior, and the relationship between coordinate and adapter distances. However, it does not establish whether clients with similar label or feature distributions use similar control mixtures, or why those mixtures are beneficial. Connecting distribution similarity, the resulting mixing weights, and prediction gains would make the proposed continuous personalization mechanism more interpretable.

Minor presentation points:

  1. In Figure 2, both the coordinate-construction block and the Bernstein-field block are titled “Client–instance coordinates.” The latter should be relabeled to distinguish the two stages.
Questions for authors

Please address the concerns raised in the weaknesses above.

Reviewer_uW5xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studied personalization via layerwise cubic Bernstein architecture that maps queries to explicit convex combination of named controls with analytically bounded sensitivity. The paper combined compact distribution coordinates, bounded instance refinement, Bernstein controls, and geometric regularization into federated learning. The paper also built partition-of-unity and Lipschitz guarantees, which is a cold-start stability/approximation bound. It then provided a supporting nonconvex local-SGD analysis of the implemented-query shared objective. The performance is evaluated by comparing with several controlled baselines.

Critical strengths

The paper studied a well-formulated problem, and attempted to address the problems during personalized federated learning. The technique part includes mathematical analysis, which are all standard steps to prove a federated learning method. Performance is evaluated with some results to validate the proposed method.

Critical weaknesses

The paper lacks strong motivations. It listed several literature, but did not highlight the problems from existing work. For example, the related work indicates the main distinction is that distribution structure indexes a smooth parameter field while instance features refine that index inside each client. Then, why do we need this for personalized federated learning? This is unclear. Similar problems exist in other literature review. Another example is the first section of related work, which says “their personalized state is still attached to a discrete client identifier”. Then, what is the problem? Is it related to privacy of client or some other problems. If this is the issue, how could this paper address the issue and validate this in its performance evaluation. Therefore, the literature review can’t explicitly highlight the existing technique challenges. The motivation of the technique part remains vague. Presentation could be significantly improved. For example, Figure 2 may be separated into multiple figures with clear corresponding explanations. The figure contains too much content with limited connections among each components. It is important to provide an overview to introduce each segment, but the architecture overview is too brief, while the figure lacks sufficient logical explanations. How could each component be connected to others? Instead of only showing four design questions, it is important to explain why each design question is necessary. Otherwise, the design does not make sense as the motivation is unclear. The proposed method seems to combine several existing methods. What are the new breakthrough and novelty? The paper should be improved to highlight the technique challenges and further link the problems to each component. Although the performance results are provided, the explanations and discussions can’t reflect the key idea to express. This makes it hard to validate the proposed method. The key points should be highlighted with the evident figures and eco the technique challenges/motivations which are introduced at the beginning.

Questions for authors

The introduction states that the unresolved question is therefore not only which model belongs to a client, but how a coherent family of models should vary across client and instance context. However, many literature discussed a series of models adjusting across client and context. Furthermore, as the personalized model belongs to each client, why this is still unresolved.

Reviewer_Kw2MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses the artificial boundary problem in personalized federated learning, where client models are either treated as entirely independent or constrained to a hard selection from a finite set of experts. The authors propose FedSPLINE, which frames personalization as a parameter field that varies continuously across a client-sample coordinate space. Each client is mapped to a 2D coordinate within a unit square plus a private, bounded offset. Within each adaptation layer, a lightweight router applies an additional, bounded sample-level displacement based on the sample's hidden state. These coordinates are used to query a tensor-product cubic Bernstein field. These adapters are combined via a convex combination and then added to the shared base weights. The backbone, coordinate encoder, router, and control points are aggregated via FedAvg, while private offsets and classification heads remain local. Evaluated on three constructed benchmarks featuring composite offsets (MNIST-C, CIFAR-10-C, and Office-Home-10) against 12 baselines and a capacity-matched MLP hypernetwork, FedSPLINE achieves the highest average accuracy (83.11%) and unlabeled cold-start accuracy (63.23%). However, the margins over the hypernetwork (0.14% for average accuracy and 0.50% for cold-start accuracy) are not statistically significant (p=0.658). Similarly, the 0.49% advantage is not significant (p=0.150), while performance in extrapolation is notably worse; furthermore, the communication overhead is approximately 2.3 times that of the baselines. Consequently, the authors regard the contribution of this work as an interpretable exploration of the trade-off between geometry and utility, rather than as a universally superior personalization paradigm.

Critical strengths
  1. I believe the problem proposed in this paper has practical significance. In particular, by reframing the question of "which model to select" in Personalized Federated Learning as "how a family of models varies continuously along the client-sample coordinate space," the paper offers a geometric interpretation.

  2. The experiments in this paper are sufficient.

Critical weaknesses
  1. My main concern with this paper is that the abstract itself appears to concede that the core contributions and experimental results are limited. Specifically, the latter part of the abstract said that the proposed method performs only 0.14%–0.19% differently from a capacity-matched MLP-based hypernetwork. Furthermore, the 0.49% advantage in interpolation is "nonsignificant" (p=0.150), while performance in extrapolation is "weaker." Finally, the method is characterized as "not a universally superior PFL strategy." In other words, the authors themselves acknowledge that their approach is indistinguishable in terms of accuracy from a baseline that replaces Bernstein weights with a standard MLP, and that it performs even worse outside the convex hull.

  2. The method fails to demonstrate a clear advantage. It outperforms the 2022 method "FedRoD" by 0.82%, yet FedRoD requires only 43% of the communication overhead used by the proposed method. I do not think that incurring significantly higher communication costs for such a marginal gain is a practical trade-off.

  3. In the ablation study (Table 7), the contribution of each component is within 1%, which is a margin smaller than the standard deviation (±1.29 to ±2.43). Moreover, the full version of the proposed method (Full FEDSPLINE) does not consistently yield the best results. This fails to substantiate the effectiveness of the proposed modules; could the authors provide an explanation?

  4. Regarding the experimental setup: the use of 12 clients, ≤420 samples per client, 24 rounds, and a single local epoch deviates significantly from the typical scale of federated learning (which generally involves a large number of clients, substantial data volumes, and multiple local epochs). The primary benchmark, MNIST-C, is custom-constructed, and there are no results reported on any standard PFL benchmarks. Additionally, in Table 1, the performance of FedLWS (matched) at 68.48% is lower than that of FedAvg (71.00%), and Feroma (matched) at 74.85% is also far below the performance of FedRep or Ditto. The figures for FedProx and FedAvg are nearly identical (rows 1 and 2). These experimental results are quite unusual.

  5. Regarding the presentation, Figures 1 and 2 contain too much information and use a font size that is too small, making them difficult to read.

Questions for authors

Please see the weakness.

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

Related papers

Discussion (0)

Sign in to comment.