acceptodds
Withdrawn from ICLR 2027

Conflict-Averse Update Decoupling for Federated Multi-Task Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Federated Multi-Task Learning (FMTL) enables clients with different tasks to collaboratively train models without sharing raw data. However, task and data heterogeneity jointly induce conflicting updates across clients. Existing FMTL methods treat the aggregated update each client receives from the server as a whole, or keep only the dimensions selected by magnitude; such updates may oppose the client's local direction, so the two cancel each other out and cause negative transfer. In this paper, we propose CAUD (Conflict-Averse Update Decoupling), a personalized aggregation method that mitigates parameter conflicts in FMTL without introducing any additional learnable parameters or network modules. Specifically, we adopt a vector obtained from a conflict-averse max-min problem as the global guidance, and decouple each client's update into consensus and client-specific parts through element-wise direction consistency with this guidance; the two parts are then modulated with closed-form weights to produce a personalized update for each client, thereby suppressing conflicts while retaining beneficial task-specific knowledge. Extensive experiments on the standard PASCAL-Context and NYUD-v2 benchmarks demonstrate that CAUD consistently improves performance across various heterogeneous client configurations.

Primary area:
optimization
Keywords:
Federated learning, Multi-Task Learning, Task set heterogeneity

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_PXJAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Paper summary The paper studies federated multi-task learning (FMTL) under task and data heterogeneity, where different clients may pursue different tasks. Such heterogeneity causes their parameter updates to have conflicting directions at the individual parameter-coordinate level. Conventional aggregation can therefore cancel useful updates and lead to negative transfer. The paper proposes Conflict-Averse Update Decoupling (CAUD) to distinguish shared, consensus information from client-specific information before aggregation.

Contributions:

  1. CAUD constructs a max-min global guidance vector within a trust region around the mean client update. The vector is designed to be compatible with all clients rather than simply averaging their potentially conflicting updates.
  2. Each client update is decomposed coordinate-wise according to whether its direction agrees with the global guidance. Consistent coordinates are treated as consensus knowledge, while conflicting coordinates are retained as client-specific knowledge.
  3. Consensus coordinates are guided by the global direction, whereas client-specific coordinates are continuously modulated according to their relative magnitude instead of being simply discarded. The resulting personalized updates require no additional learnable parameters or modules.

Validation:

The method is evaluated on PASCAL-Context and NYUD-v2 under heterogeneous FMTL settings and compared with a broad range of federated and multi-task aggregation methods. CAUD achieves a 4.18% Δm, compared with 2.18% for the strongest baseline, FedHCA2, in the main heterogeneous setting.

Critical strengths
  1. The key strength is the explicit focus on parameter-level update conflicts rather than treating each client update as an indivisible vector. By decoupling consensus and client-specific coordinates and handling them differently, CAUD provides a more fine-grained way to balance shared and task-specific knowledge. This is consequential because parameter conflict is a fundamental challenge in heterogeneous FMTL, and the proposed formulation directly targets the mechanism that can cause update cancellation and negative transfer.
  2. CAUD achieves a substantially higher overall improvement than the strongest baseline in the main heterogeneous setting (Δm of 4.18% versus 2.18% for FedHCA2), while ablations show progressive gains from conflict-averse guidance, coordinate-level decoupling, and adaptive modulation. The parameter-level analysis further supports the claim that CAUD produces more coordinated cross-client updates. This combination of comparative, ablation, and mechanistic evidence makes the empirical conclusions more convincing than relying on overall performance alone.
Critical weaknesses
  1. The proposed global guidance vector is explicitly formulated as the update-space analogue of CAGrad, and the paper acknowledges that the same max-min conflict-averse formulation has already been federated in prior work. The main additional contribution is therefore the subsequent sign-based coordinate decoupling and scalar modulation. However, the decoupling itself relies on a relatively simple heuristic, which is that the coordinates whose update signs agree with the global guidance are labeled as consensus, while the remaining coordinates are treated as client-specific. This binary sign criterion does not establish whether a conflicting coordinate is genuinely task-specific or merely an undesirable optimization conflict. Consequently, the conceptual gap between CAUD and prior conflict-averse aggregation/personalization methods is not sufficiently substantial to support the paper's broad claim of a new solution to parameter-level conflict in FMTL.
  2. The central design of CAUD depends on interpreting directionally consistent coordinates as shared knowledge and conflicting coordinates as client-specific knowledge. Yet the experimental evidence mainly demonstrates that the resulting update transformation improves task performance and increases cross-client update similarity, neither directly verifies the semantic or functional meaning of the two coordinate groups. In fact, the ablation shows that simply discarding the client-specific coordinates (3.86%) performs better than retaining them without modulation (3.60%), while the proposed adaptive modulation provides only a further 0.32-point gain (4.18%). This makes it difficult to establish that the conflicting coordinates actually contain useful task-specific information, rather than being predominantly harmful interference that can simply be suppressed. Since this assumption is fundamental to the claimed “consensus vs. client-specific knowledge” interpretation, the lack of direct validation substantially weakens the methodological justification of CAUD.
Questions for authors
  1. What is the empirical or theoretical evidence that sign consistency with the global guidance vector reliably identifies “consensus” knowledge, while sign inconsistency corresponds to “client-specific” knowledge?
  2. Since βᵢ is a single scalar applied uniformly to all client-specific coordinates, how does CAUD distinguish useful task-specific updates from harmful conflicting updates at the coordinate level?
Reviewer_SnviOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes CAUD, which mitigates update conflicts in heterogeneous federated multi-task learning by employing a conflict-avoidance global guidance vector. The method partitions client updates into shared and personalized components based on sign consistency, and then combines them via closed-form weights. The approach is concise and does not introduce additional learnable modules. It achieves strong overall performance across various client configurations, and ablation studies validate the effectiveness of coordinate-level decoupling. However, the primary novelty lies mainly in the combination and application of existing mechanisms. Treating sign consistency as a proxy for knowledge sharing remains overly simplistic, and the theoretical analysis has not been directly linked to actual training objectives. Overall, this is a practically valuable improvement to personalized aggregation, but there is still room to strengthen the method's novelty and the rigor of the experimental validation.

Critical strengths

S1. The paper has a well-defined problem scope, focusing on the fine-grained update conflicts in heterogeneous federated multi-task learning.

S2. The update decoupling mechanism is concise, effectively balancing shared information and client-specific characteristics. The proposed CAUD partitions the coordinate space based on the sign consistency between local updates and global guidance.

S3. The experiments cover a variety of client configurations, and the ablation studies clearly support the core design.

Critical weaknesses

W1. The biggest concern is the novelty of this paper. The global guidance for conflict avoidance builds upon the principles of CAGrad and existing federated methods such as FedHCA2, and the sign masking is also inspired by model merging literature. Therefore, the primary innovation of this paper lies in integrating these mechanisms into a personalized update rule.

W2. Treating sign consistency as a proxy for knowledge sharability remains overly simplistic. Updates with the same direction do not necessarily imply that sharing is beneficial, nor do opposite directions inevitably lead to negative transfer. Furthermore, updates close to zero may cause the mask to switch frequently due to noise.

W3. The theoretical derivation targets artificially defined auxiliary objectives, providing limited guarantees for actual training performance. Furthermore, the current analysis primarily serves as a rationale for the design, and is insufficient to support conclusions regarding overall convergence or controlled negative transfer.

Questions for authors

See weaknesses.

Reviewer_n6L5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CAUD to address conflicts among client parameter updates in federated multi-task learning, where directly averaging updates can lead to negative transfer. CAUD first constructs a global guidance direction that seeks to accommodate all clients. It then separates each client’s update into a shareable consensus component and a client-specific component according to their directional agreement, and adaptively shrinks the conflicting component. The method requires no additional trainable modules and achieves strong overall performance on two dense prediction datasets.

Critical strengths
  1. The paper directly addresses directional conflicts among client parameter updates. By separating shareable components from client-specific components that should be preserved, it tackles negative transfer in federated multi-task learning.
  2. CAUD generates personalized updates during aggregation without adding trainable modules to the model, making it relatively easy to integrate with existing models.
  3. The authors evaluate the method on two datasets under different client task configurations and conduct ablation studies, providing reasonably thorough experimental support.
Critical weaknesses
  1. Prior work such as ConFREE and FedCALM also addresses conflicts among client updates during server aggregation by constructing conflict-averse directions or selectively retaining updates. Although these methods primarily study personalized federated learning rather than the federated multi-task setting considered here, their core ideas are closely related. The authors do not adequately explain how CAUD differs from these methods or provide corresponding comparisons.

  2. The main table reports task-level metrics, some of which correspond to individual single-task clients, but it does not clearly show the overall benefit for each client or how those benefits are distributed. Since the method generates a distinct personalized update for each client, the authors should report per-client performance, worst-client performance, and the number of clients experiencing negative transfer. Average metrics alone may conceal performance degradation for individual clients.

  3. The main experiments involve only six clients and focus on two visual dense prediction datasets. The method has not been evaluated with larger numbers of clients, real cross-institutional distribution shifts, different model architectures, or non-visual tasks. Its empirical validation is therefore limited.

  4. The expression for the guidance vector contains Uw⋆/∥Uw⋆∥U_{w^\star}/\|U_{w^\star}\|. This term is undefined when the optimally weighted update Uw⋆=0U_{w^\star}=0, a case that can arise from opposing client updates. The authors should explain how Eq. (4) is solved in this case and specify how it is handled in the implementation.

  5. The optimal weight derived in Appendix Eqs. (14)–(16) is ∥Pi∥/(∥Pi∥+∥uP∥)\|P_i\|/(\|P_i\|+\|u_P\|), whereas the implemented weight in Eq. (10) includes an additional ϵ\epsilon in the denominator. The stabilizing term is understandable, but Eq. (10) is no longer exactly the unique minimizer of the stated objective. The authors should qualify this claim or provide a derivation consistent with the implemented formula.

Questions for authors

See Weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.