Pith. sign in

REVIEW 4 major objections 3 minor 45 references

FedWeave: Rethinking the Unit of Specialization in Heterogeneous Federated MoE-LoRA

T0 review · 4 major / 3 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Federated LLM adaptation improves when LoRA experts are aggregated over latent patterns while the router trains over whole clients.

desk verdict Genuinely new aggregation granularity for federated MoE-LoRA, backed by honest ablations but a theory that presupposes the main premise and experiments confined to a high-purity regime. read the letter →

arxiv 2607.26618 v1 pith:BUT4KPGA submitted 2026-07-29 cs.LG cs.CL

Donghang Duan , Xu Zheng , Lizong Zhang , Chong Mu , Meng Han This is my paper · ORCID
classification cs.LGcs.CL
keywords federatedlearningLoRAmixtureofexpertstaskheterogeneityasymmetricaggregationprototypediscoverysparseinferenceparameter-efficientfine-tuning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Federated fine-tuning of large language models often fails when a single client's data mixes several latent instruction patterns: averaging them into one client-level update entangles adaptation signals before they reach the server. FedWeave argues that the two parts of a mixture-of-experts LoRA system need different aggregation units—experts require pattern-pure updates from aligned within-client prototypes, while the router requires contrast from complete client trajectories spanning all patterns. The paper formalizes this purity–contrast principle with convergence bounds: expert drift is controlled by routed off-pattern contamination, router error separates interleaving-order bias from reset-consensus error, and sparse Top-1 inference risk is bounded by router suboptimality. On a controlled multi-task benchmark with two mainstream 2-3B backbones, FedWeave improves macro-average scores over the evaluated client-level baselines across three heterogeneity levels, and sparse inference with one active expert preserves nearly all soft-routing quality.

What carries the argument

The purity–contrast principle, operationalized as asymmetric aggregation: expert updates are weighted by matched-prototype sample counts and aggregated over aligned local buckets, while the single global router is updated along each client's complete bucket trajectory. Prototype discovery uses k-means on frozen-backbone embeddings with silhouette-based count selection; cross-client alignment uses layer-wise cosine distances between LoRA-B (zero-initialized low-rank adapter factor) warmup signatures, clustered agglomeratively. The analysis rests on three results: Theorem 1's contamination-controlled expert stationarity bound (O(1/U) convergence to a neighborhood set by off-pattern routed weig

What would settle it

Construct a federated benchmark with deliberately overlapping or drifting task patterns so that k-means buckets on frozen embeddings and LoRA-B signature alignment have near-random purity (well below the 0.95 global purity reported here), then run FedWeave against a client-level MoE-LoRA baseline under equal capacity; if the macro-average advantage vanishes or reverses, the separability assumption is confirmed as load-bearing, whereas if FedWeave still wins, purity-contrast is not the active mechanism in that regime.

Watch

Extended reading notes

Core claim

The paper's central claim is that aggregation granularity should be asymmetric in federated MoE-LoRA: no single granularity serves both components. LoRA experts should be aggregated over aligned local prototypes—clusters of examples within each client that share a latent instruction pattern—because prototype-level updates preserve the pattern coherence needed for specialization and prevent cross-pattern gradient cancellation. The router, by contrast, should be trained over each client's complete interleaved bucket trajectory and aggregated as one client-level delta, because routing decisions need cross-pattern contrast to compare experts. FedWeave discovers prototypes by k-means on frozen-ba

Load-bearing premise

The load-bearing premise, stated in the paper's Limitations, is that local latent instruction patterns are sufficiently separable in frozen-backbone embeddings and in short LoRA-B warmup signatures for prototype recovery and cross-client alignment to succeed; with weak separability or drifting patterns, bucket assignments become noisy and prototype-level expert aggregation loses its advantage over client-level aggregation.

Editorial extensions

If this is right

  • Aggregation granularity becomes a first-class design choice: prototype-level expert groups can replace client-level expert units without changing total LoRA capacity.
  • Persistent router training over whole client trajectories avoids the consensus-error residual of per-bucket router resets, leaving only a schedule-order bias that vanishes under randomized bucket sampling.
  • Sparse inference with a single active expert retains nearly all soft-routing quality—roughly a 0.001 macro-score gap while cutting mean latency by about a third in the reported setting.
  • Specialization emerges without task labels: patterns are recovered from frozen representations and LoRA-B warmup signatures, and the router learns distinct dominant experts for each task.
  • The benefit is consistent across two mainstream 2-3B backbones and three Dirichlet heterogeneity levels, with the largest margin at intermediate heterogeneity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the purity–contrast principle generalizes, other federated systems with specialized parameters plus a shared controller—such as personalized heads with a global feature extractor—could benefit from the same split: specialize where patterns are coherent, keep the controller on mixed trajectories.
  • The paper's own Limitations bound the conclusion to 2B–3B backbones and controlled four-task mixtures; natural, drifting instruction streams and larger models remain untested, so the principle's practical ceiling is open.
  • A testable extension is to add routing-entropy regularization that concentrates routed weight on the target expert; Theorem 1's contamination term would shrink directly, likely accelerating expert convergence under the same analysis.
  • Because alignment relies on one-time LoRA-B signatures, online alignment or re-bucketing after drift would be needed for nonstationary clients; the method's benefit likely degrades as pattern separability weakens.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. This paper proposes FedWeave, a federated MoE-LoRA method based on asymmetric aggregation granularity: local clients first discover within-client pattern buckets via frozen-backbone embeddings plus k-means, compute LoRA-B warmup signatures, and the server aligns buckets across clients into global expert groups. Expert updates are aggregated at the aligned-prototype level, while a single global router is updated over each client's complete interleaved bucket trajectory and aggregated client-wise. At inference only the Top-1 expert is activated. The paper claims a purity–contrast principle, proves (i) expert convergence to a stationarity neighborhood controlled by off-target routing contamination, (ii) PL-type router convergence in which deterministic interleaving incurs an order-bias residual and reset-and-average incurs a consensus residual, and (iii) a gap-calibration bound transferring router suboptimality to Top-1 risk. Experiments on CoEdIT, GSM8K, TweetEval, and ARC-C with Llama3.2-3B and Gemma-2-2B report consistent macro-average gains over FedIT, FFA-LoRA, FedSA-LoRA, and FedLEASE (e.g., 0.5872 vs 0.5649 on Llama3.2-3B), with ablations on aggregation granularity, clusterers, and sparse inference.

Significance. The purity–contrast principle is a genuinely useful design lens for federated MoE-LoRA: it separates the aggregation granularity for experts (pattern-coherent prototypes) from that for the router (complete client trajectories). If the central comparison were established unconditionally, this would be a meaningful step beyond client-level specialization. The manuscript has notable strengths: a reproducible code package with data manifests, a controlled benchmark with matched capacity, clear ablations, transparent limitations, and an honest theory that states its conditions rather than hiding them. However, the main theoretical advantage is conditional on quantities the method is designed to improve, and the empirical evidence is confined to a deliberately high-purity regime. The significance is therefore moderate: the idea is promising and the empirical results are positive, but the load-bearing claims need additional support before the paper can be accepted.

major comments (4)
  1. [Appendix C, Corollary 1] The expert-side advantage is conditional on χ_m^u ≤ α_client,u_m (Corollary 2), but this is the very property the method is intended to deliver. Lemma 1 decomposes χ_m into ε_loc + ε_align + ε_leak, yet no theorem bounds these quantities from data or shows they are smaller than the client-level analogue; Eq. (22) merely re-expresses the contamination. Thus the convergence theorem does not independently establish prototype-level purity. Please add a formal sufficient condition (e.g., on cluster purity and alignment error) under which χ_m ≤ α_client, or empirically measure χ_m and α_client from the trained routers; otherwise the central claim rests on an assumption.
  2. [Experiments / Limitations] The empirical support covers only the favorable regime. The paper states the benchmark "contains distinct instruction patterns that can be recovered from the frozen representations," and Table 3 reports global purity 0.9515; the Limitations concedes that FedWeave assumes sufficient separability. No experiment varies pattern separability, signature noise, or drift, so the 0.0223 gain over w/o prototype-expert (Table 2) is demonstrated only at high purity. I recommend a controlled degradation study—e.g., injecting noise into embeddings/signatures or partially mixing latent patterns—and reporting macro score versus purity, to show whether the advantage degrades gracefully.
  3. [Appendix C, Theorems 4–5] The router comparison is not resolved by the bounds. Persistent training yields E_i,S + O_i,S, reset-and-average yields E_i,S + C_i,S, and the text states "their magnitudes determine the bound ordering." Since deterministic interleaving makes O_i,S nonzero, the statement that persistent trajectories "replace the reset consensus residual" is not proven; it only identifies a different residual. Please provide a condition on bucket heterogeneity/order bias under which O_i,S ≤ C_i,S, or an empirical decomposition of these terms in the actual training schedule; without this, the router-contrast half of the asymmetry is not theoretically established.
  4. [Table 2 / Eq. (40)] The w/o client-router ablation changes two things simultaneously: it fragments the router state and it reweights endpoints (effective weights eπ_c ∝ a_c H_c, Eq. (40)). The theory in Appendix C shows weight mismatch enters the same residual as b_wt, so the 0.0294 loss gap cannot be attributed solely to loss of cross-bucket contrast. An ablation that preserves endpoint weights while resetting router states, or one that holds state continuity while inducing weight mismatch, would disentangle these factors.
minor comments (3)
  1. [Appendix C, Eq. (15)] The [·]_+ notation is used before it is defined; please define it explicitly at first use.
  2. [Theorem 1] The text says "O(1/U) convergence," but the bound has a non-vanishing contamination/noise floor. Please state explicitly that the first term is O(1/U) only when the effective budget A_m,U grows linearly, and that the remaining terms are a stationary neighborhood.
  3. [Appendix A, Table 4] The hardware row lists "48 GB GPU memory; 49,140 MiB" for an RTX 4090. The standard RTX 4090 has 24 GB; this appears to be a typo or requires clarification.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: theoretical comparisons are explicitly conditional and empirical claims are tested against external baselines.

full rationale

FedWeave's derivation chain is not circular. The claimed theoretical advantage of prototype-level expert aggregation is an explicitly conditional comparison: the prototype bound is tighter only if the routed off-pattern contamination χ_m is no larger than the client-level contamination α_client. The paper states this condition openly ('the prototype bound is no looser if χm(θu_m) ≤ α_client,u_m at every active step') and does not present it as a derived fact. Similarly, the router analysis identifies O_i,S and C_i,S residuals but explicitly notes that 'their magnitudes determine the bound ordering,' so it does not assert an unproven superiority. The sparse-risk bound is standard gap calibration. The central empirical claim is validated against external baselines on a controlled benchmark, and the paper's Limitations section concedes the recoverability assumption ('FedWeave assumes that local patterns are sufficiently separable in representation and adaptation-signature space for prototype recovery and cross-client alignment'). This limits generalizability but does not reduce any prediction to a fitted input or to a self-citation. No load-bearing self-citations or imported uniqueness theorems appear; related-work citations such as FedTreeLoRA are not used to justify FedWeave's core mechanism.

Assumptions & free parameters 2 free parameters · 7 assumptions · 0 invented entities

The central claim rests on pattern-separability assumptions and unverified router conditions; no new physical entities are introduced. The silhouette-selected expert count is data-adaptive model selection, not a parameter fitted to the target; the router/optimizer hyperparameters are conventional.

free parameters (2)
  • Per-client bucket count Ci = Silhouette-selected in 2..min(8, n_i-1); not reported per client
    The number of local prototypes is chosen by silhouette score on k-means clustering of frozen embeddings, making it data-adaptive; the central method depends on this choice but the paper evaluates only aggregate alignment quality.
  • Global expert count M* = 4 in reported main runs for both backbones
    Selected via silhouette over agglomerative clustering of bucket signatures; the method adapts it, and the main runs used four experts to match rank-32 single-adapter capacity. Not fitted to the test target.
assumptions (7)
  • domain assumption Latent instruction patterns are recoverable as k-means clusters in frozen-backbone embeddings
    FedWeave clusters each client's frozen embeddings into buckets; if the patterns are not separated, bucket purity is low and expert updates are contaminated. High purity in Table 3 holds only for the controlled benchmark.
  • domain assumption Adaptation signatures from 10-step LoRA-B warmups identify which buckets can share an expert across clients
    Cross-client alignment (Eq. 1-2) uses LoRA-B signatures; this assumes short warmup training produces stable, task-discriminative B factors.
  • standard math Assumption 1: bounded routed gradients and within-pattern shift (Eq. 17)
    Used in Proposition 1 and Theorem 1 to bound contamination.
  • standard math Assumption 2: LE-smooth target objective, bounded conditional variance of routed stochastic direction
    Standard SGD conditions for Theorem 1.
  • domain assumption Assumption 3: LR-smooth bucket router losses and mu-PL condition on the client router objective
    PL condition is not guaranteed for neural router losses; used in Theorems 2, 4, 5.
  • domain assumption Assumption 4: gap calibration c_gap * S_i(φ) ≤ R_i(φ) - R_oracle + β_mix
    Needed for sparse-inference risk bound; unverified for LLM losses except in a linear-mixture toy case.
  • domain assumption Clean regime K = M* with a bijection between latent patterns and experts for oracle-routing statements
    The one-to-one oracle-routing statements in the theory require an exact match between patterns and experts; in practice K is unknown and M*≈4.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FedWeave: Rethinking the Unit of Specialization in Heterogeneous Federated MoE-LoRA." pith.science (2026). https://pith.science/paper/BUT4KPGA

@misc{pith2026260726618,
  author       = {Pith},
  title        = {Pith review of: FedWeave: Rethinking the Unit of Specialization in Heterogeneous Federated MoE-LoRA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BUT4KPGA}},
  note         = {Machine review of arXiv:2607.26618}
}
read the original abstract

Federated PEFT enables LLMs to collaboratively adapt to decentralized private data without sharing raw examples. However, task heterogeneity across clients can cause cross-task interference and gradient conflicts during aggregation. Federated MoE-LoRA addresses this challenge through specialized LoRA experts and conditional routing. Yet existing methods typically specialize at client granularity, implicitly assuming task-coherent clients. Our core insight is that experts need purity, namely pattern-coherent updates that preserve specialization, whereas routers need contrast, namely mixed-task observations that support expert comparison. We propose FedWeave, a framework that adopts asymmetric aggregation, separating expert aggregation from router optimization to meet these two requirements. FedWeave uses unsupervised prototype discovery to form local buckets and align them across clients, enabling prototype-level expert aggregation while retaining mixed-task client trajectories for router training. At inference, FedWeave performs sparse inference with one active expert while preserving nearly all soft-routing performance. Our theoretical analysis explains why asymmetric aggregation is advantageous: it controls expert convergence in stationarity through off-pattern contamination, identifies the consensus error induced by fragmented router trajectories, and bounds sparse-inference risk. On a heterogeneous multi-task benchmark with mainstream LLM backbones, FedWeave consistently outperforms strong baselines, while ablations verify the effectiveness of our design.

Figures

Figures reproduced from arXiv: 2607.26618 by the authors.

Figure 1
Figure 1. FedWeave overview. Aligned local prototypes provide pattern-coherent batch supervision for experts, whereas the single global router retains cross-pattern contrast by training over each client’s complete bucket trajectory. The server aggregates the two components at different granularities. {Bi,c} Ci c=1. We call the conceptual specialization unit a pro￾totype and its concrete local implementation a bucket. This exp… view at source ↗
Figure 3
Figure 3. Macro-average score and teacher-forced loss on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Sparse-inference quality–latency trade-off over [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 5 linked inside Pith

  1. [1]

    2020 , publisher =

    Barbieri, Francesco and Camacho-Collados, Jose and Espinosa Anke, Luis and Neves, Leonardo , booktitle =. 2020 , publisher =

    2020

  2. [2]

    Think You Have Solved Question Answering? Try

    Clark, Peter and Cowhey, Isaac and Etzioni, Oren and Khot, Tushar and Sabharwal, Ashish and Schoenick, Carissa and Tafjord, Oyvind , journal =. Think You Have Solved Question Answering? Try

  3. [3]

    arXiv preprint arXiv:2110.14168 , year =

    Training Verifiers to Solve Math Word Problems , author =. arXiv preprint arXiv:2110.14168 , year =

    arXiv

  4. [4]

    Dettmers, Tim and Pagnoni, Artidoro and Holtzman, Ari and Zettlemoyer, Luke , booktitle =

  5. [5]

    Journal of Machine Learning Research , volume =

    Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity , author =. Journal of Machine Learning Research , volume =

  6. [6]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

    Towards Instance-adaptive Inference for Federated Learning , author =. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages =

  7. [7]

    Advances in Neural Information Processing Systems , volume =

    An Efficient Framework for Clustered Federated Learning , author =. Advances in Neural Information Processing Systems , volume =

  8. [8]

    International Conference on Learning Representations , year =

    Selective Aggregation for Low-Rank Adaptation in Federated Learning , author =. International Conference on Learning Representations , year =

Show all 45 references
  1. [9]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle =

  2. [10]

    Foundations and Trends in Machine Learning , volume =

    Advances and Open Problems in Federated Learning , author =. Foundations and Trends in Machine Learning , volume =. 2021 , doi =

    2021

  3. [11]

    2020 , publisher =

    Karimireddy, Sai Praneeth and Kale, Satyen and Mohri, Mehryar and Reddi, Sashank and Stich, Sebastian and Suresh, Ananda Theertha , booktitle =. 2020 , publisher =

    2020

  4. [12]

    Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak--

    Karimi, Hamed and Nutini, Julie and Schmidt, Mark , booktitle =. Linear Convergence of Gradient and Proximal-Gradient Methods Under the Polyak--. 2016 , publisher =

    2016

  5. [13]

    Proceedings of Machine Learning and Systems , volume =

    Federated Optimization in Heterogeneous Networks , author =. Proceedings of Machine Learning and Systems , volume =

  6. [14]

    Liao, Mengqi and Chen, Wei and Shen, Junfeng and Guo, Shengnan and Wan, Huaiyu , booktitle =

  7. [15]

    Advances in Neural Information Processing Systems , volume =

    Federated Multi-Task Learning under a Mixture of Distributions , author =. Advances in Neural Information Processing Systems , volume =

  8. [16]

    Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , series =

    Communication-Efficient Learning of Deep Networks from Decentralized Data , author =. Proceedings of the 20th International Conference on Artificial Intelligence and Statistics , series =. 2017 , publisher =

    2017

  9. [17]

    arXiv preprint arXiv:2408.00118 , year =

    arXiv

  10. [18]

    Personalized Federated Fine-tuning for Heterogeneous Data: An Automatic Rank Learning Approach via Two-Level

    Hao, Jie and Wu, Yuman and Payani, Ali and Lee, Myungjin and Liu, Mingrui , journal =. Personalized Federated Fine-tuning for Heterogeneous Data: An Automatic Rank Learning Approach via Two-Level

  11. [19]

    2023 , publisher =

    Raheja, Vipul and Kumar, Dhruv and Koo, Ryan and Kang, Dongyeop , booktitle =. 2023 , publisher =

    2023

  12. [20]

    arXiv preprint arXiv:2107.06724 , year =

    Federated Mixture of Experts , author =. arXiv preprint arXiv:2107.06724 , year =

    arXiv

  13. [21]

    Journal of Computational and Applied Mathematics , volume =

    Silhouettes: A Graphical Aid to the Interpretation and Validation of Cluster Analysis , author =. Journal of Computational and Applied Mathematics , volume =. 1987 , doi =

    1987

  14. [22]

    IEEE Transactions on Neural Networks and Learning Systems , volume =

    Clustered Federated Learning: Model-Agnostic Distributed Multi-Task Optimization under Privacy Constraints , author =. IEEE Transactions on Neural Networks and Learning Systems , volume =. 2021 , doi =

    2021

  15. [23]

    International Conference on Learning Representations , year =

    Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer , author =. International Conference on Learning Representations , year =

  16. [24]

    Shen, Zhikang and Lu, Jianrong and Wan, Haiyuan and Chen, Jianhai , journal =

  17. [25]

    Improving

    Sun, Youbang and Li, Zitao and Li, Yaliang and Ding, Bolin , booktitle =. Improving

  18. [26]

    arXiv preprint arXiv:2604.00050 , year =

    Task-Centric Personalized Federated Fine-Tuning of Language Models , author =. arXiv preprint arXiv:2604.00050 , year =

    arXiv

  19. [27]

    2026 , doi =

    Bian, Jieming and Wang, Lei and Zhang, Letian and Xu, Jie , journal =. 2026 , doi =

    2026

  20. [28]

    Bian, Jieming and Wang, Lei and Zhang, Letian and Xu, Jie , journal =

  21. [29]

    Heterogeneous

    Cho, Yae Jee and Liu, Luyang and Xu, Zheng and Fahrezi, Aldi and Joshi, Gauri , booktitle =. Heterogeneous. 2024 , publisher =

    2024

  22. [30]

    Lee, Seanie and Park, Sangwoo and Lee, Dong Bok and Wagner, Dominik and Seong, Haebin and Bocklet, Tobias and Lee, Juho and Hwang, Sung Ju , booktitle =

  23. [31]

    Adaptive

    Wang, Lei and Bian, Jieming and Zhang, Letian and Xu, Jie , booktitle =. Adaptive

  24. [32]

    2024 , doi =

    Wang, Ziyao and Shen, Zheyu and He, Yexiao and Sun, Guoheng and Wang, Hongyi and Lyu, Lingjuan and Li, Ang , booktitle =. 2024 , doi =

    2024

  25. [33]

    arXiv preprint arXiv:2010.02056 , year =

    Specialized Federated Learning using a Mixture of Experts , author =. arXiv preprint arXiv:2010.02056 , year =

    2010 arXiv

  26. [34]

    Towards Building the Federated

    Zhang, Jianyi and Vahidian, Saeed and Kuo, Martin and Li, Chunyuan and Zhang, Ruiyi and Yu, Tong and Zhou, Yufan and Wang, Guoyin and Chen, Yiran , journal =. Towards Building the Federated

  27. [35]

    Mei, Hanzi and Cai, Dongqi and Zhou, Ao and Wang, Shangguang and Xu, Mengwei , journal =

  28. [36]

    Chen, Shaoxiang and Jie, Zequn and Ma, Lin , journal =

  29. [37]

    Hybrid Routing for a Mixture of

    Huang, Yitong and Yang, Ziqi and Wang, Zihui and Qi, Jianzhong and Yu, Rongshan and Fan, Xiaoliang and Wang, Cheng , journal =. Hybrid Routing for a Mixture of. 2026 , doi =

    2026

  30. [38]

    Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages =

    Federated Low-Rank Adaptation for Foundation Models: A Survey , author =. Proceedings of the Thirty-Fourth International Joint Conference on Artificial Intelligence , pages =. 2025 , doi =

    2025

  31. [39]

    Bian, Jieming and Wang, Lei and Zhang, Letian and Xu, Jie , booktitle =

  32. [40]

    2025 , doi =

    Fan, Boyu and Su, Xiang and Tarkoma, Sasu and Hui, Pan , journal =. 2025 , doi =

    2025

  33. [41]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    Towards Robust and Efficient Federated Low-Rank Adaptation with Heterogeneous Clients , author =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2025 , doi =

    2025

  34. [42]

    and Vepakomma, Praneeth , journal =

    Singhal, Raghav and Ponkshe, Kaustubh and Vartak, Rohit and Varshney, Lav R. and Vepakomma, Praneeth , journal =. 2026 , url =

    2026

  35. [43]

    Beyond Factor Aggregation: Gauge-Aware Low-Rank Server Representations for Federated

    Chen, Jinqian and Liu, Chang and Zhu, Jihua , journal =. Beyond Factor Aggregation: Gauge-Aware Low-Rank Server Representations for Federated

  36. [44]

    Yang, Nuocheng and Wang, Sihua and Huan, Ouwen and Chen, Mingzhe and Quek, Tony Q. S. and Yin, Changchuan , journal =. Wireless Federated Multi-Task

  37. [45]

    Neurocomputing , volume =

    A Communication-Efficient Personalized Federated Learning Framework Driven by Parameter Decoupling , author =. Neurocomputing , volume =. 2026 , doi =

    2026

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.