acceptodds
Under review as a conference paper at ICLR 2027

Bi-granularity Weighted Aggregation for Asynchronous Federated Learning

Abstract

Asynchronous federated learning (AFL) enables collaborative model training under heterogeneous computation and communication latency, where participating clients may update and transmit their local models at substantially different speeds. Existing AFL methods typically discard updates from stragglers or reduce their aggregation weights according to update staleness. However, under severe system heterogeneity, such strategies may underutilise valuable client information, slow convergence, and increase discrepancies between the aggregated global model and the models that would have been obtained with more complete client participation. To address these challenges, we propose a novel bi-granularity weighted aggregation framework for AFL. The key idea is to reconstruct a pseudo local model for each straggling client, allowing federated aggregation to proceed by jointly incorporating received local models and reconstructed models for unavailable stragglers. To balance reconstruction quality and computational efficiency, rather than generating an entire model from scratch, our method reconstructs each pseudo model by reassembling neural blocks selected from available client models and leveraging the straggler's statistical fingerprints. During global aggregation, client contributions are jointly weighted at both the model level and neural-block level, enabling the server to account for heterogeneity at different granularities. Extensive experiments across diverse large-scale asynchronous settings demonstrate that the proposed method consistently improves performance over representative AFL baselines. Further analysis shows that the reconstructed pseudo models closely approximate the actual local models subsequently obtained from straggling clients, providing empirical evidence for the effectiveness of the proposed reconstruction mechanism.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Machine learning, Distributed machine learning, Privacy-Preserving machine learning, Federated learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZqK6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a bi-granularity weighted aggregation framework for asynchronous federated learning (AFL). The main idea is to reconstruct pseudo local models for straggling clients using neural blocks from already received client models, allowing their estimated contributions to participate in global aggregation without waiting for their actual updates.

The framework first obtains client-specific activation fingerprints through a one-time warm-up. During asynchronous training, it groups available neural blocks using representation similarity, constructs candidate models through ordered block borrowing, and selects pseudo models through fingerprint-based matching. The reconstructed models are then aggregated together with received local models.

Experiments on SVHN, CIFAR-10, and CIFAR-100 demonstrate competitive performance against several AFL baselines under different levels of data and system heterogeneity.

Critical strengths

1. An interesting approach to straggler compensation in AFL. Rather than simply discarding or downweighting delayed updates, the paper attempts to incorporate estimated contributions from unavailable clients through model reassembly. This provides an interesting perspective on handling stragglers in asynchronous federated learning.

2. Relatively comprehensive experimental evaluation. The paper evaluates different datasets, statistical heterogeneity levels, and asynchronous settings. It also includes sensitivity analyses and ablation studies for candidate construction, block borrowing, source diversity, and fingerprint-based matching.

Critical weaknesses

1. Unclear distinction between model reassembly and block-wise reweighting.

Model reassembly has already been explored in federated learning, such as pFedHR (NeurIPS 2023) and pFedClub (NeurIPS 2024). While this paper applies model reassembly to approximate unavailable straggler models in AFL, the resulting aggregation can also be interpreted as a form of block-wise client reweighting. Without candidate calibration, aggregating reconstructed models is mathematically equivalent to reweighting the available client models at different block positions. Although calibration prevents strict equivalence for the complete method, the authors should clarify the distinction and compare against a simpler block-wise reweighting baseline.

2. Reliability of static client fingerprints throughout asynchronous training.

The method relies on client fingerprints obtained during a one-time warm-up and reused throughout asynchronous training. However, as client models evolve, these fixed representations may no longer accurately reflect their current characteristics. Although the paper evaluates different warm-up epochs, it does not directly investigate representation drift or its impact on pseudo-model matching. Evaluating fingerprint stability over time and comparing static fingerprints with periodically refreshed ones would help validate this assumption.

3. Unclear handling of late-arriving genuine client updates.

The proposed framework incorporates reconstructed pseudo models for straggling clients before their genuine updates arrive. However, the paper does not clearly explain how these genuine updates are handled afterward. If they are aggregated again, the same client training task may contribute more than once. The authors should clarify whether late-arriving updates are discarded, incorporated again, or used for correction, and explain how duplicate contributions are addressed.

4. Lack of theoretical analysis of the proposed aggregation mechanism.

The paper mainly relies on empirical evaluation without providing theoretical analysis of how reconstructed pseudo models affect asynchronous optimization. In particular, reconstruction errors may introduce additional bias into global aggregation, especially under data heterogeneity and stale updates. A theoretical analysis relating reconstruction error to aggregation bias or convergence behavior would help explain under what conditions the proposed method improves upon aggregation using only received client updates.

Questions for authors
  1. Since the pseudo models are constructed from neural blocks of already received client models, to what extent can the proposed aggregation be interpreted as block-wise client reweighting? What benefits does model reassembly provide beyond a simpler reweighting strategy?

  2. How stable are the client fingerprints obtained during one-time warm-up? Have the authors investigated representation drift during asynchronous training and its effect on pseudo-model matching?

  3. After a pseudo model for a straggling client has been included in aggregation, how is the corresponding genuine local model handled when it eventually arrives? Can the same client training task contribute more than once?

  4. Can the authors provide theoretical justification for incorporating reconstructed models into asynchronous aggregation, particularly regarding how reconstruction error affects global update bias and convergence?

3 comments
AuthorsOpenReview ↗

We clarify the distinction between model reassembly and block-wise reweighting, present fingerprint stability visualizations, and explain how our buffer-free framework handles late-arriving updates.

W1&Q1: Model Reassembly vs. Block-wise Reweighting. We thank the reviewer for noting the link between model reassembly and block-wise reweighting. The key distinction from direct block-wise reweighting lies in how we select block combinations and adapt the assembled model. For each straggling client, we use its cache client fingerprint to evaluate and select complete candidate models. Block selection therefore depends on the overall behavior of the assembled model. We then fine-tune the selected candidate on proxy data to jointly adapt blocks from different sources. Our full pipeline therefore combines history-guided block selection with candidate-specific parameter adaptation, providing an optimization mechanism beyond direct block-wise reweighting.

W2&Q2: Fingerprint Stability. Because the fingerprints cached during one-time warm-up are reused to match reconstructed models to straggling clients, we provide additional visualizations to examine fingerprint stability. In APPENDIX E, we provide two visualizations of activation patterns from the same client model at three warm-up settings: Step 0, Default, and Round 40. At each step, we use four fixed proxy samples and extract activations from the same feature layer, maintaining consistent channel indices and fixed per-channel color scales. The first visualization displays 64 selected channels per proxy sample to facilitate inspection of individual responses. The second displays all 512 channels per proxy sample and highlights six locations, with corresponding enlarged patches shown below each panel. At Step 0, the activation maps are generally weak under the shared color scales. Distinct spatial responses emerge under the default setting. Between Default and Round 40, the highlighted channels exhibit similar spatial response patterns, although activation intensities and local details continue to change. These visualizations provide qualitative evidence of pattern persistence. To ensure that we address the concern precisely, could the reviewer explain which representations are being considered, during which stage of training the drift occurs, and what form of representational change is of concern? We would appreciate further clarification on these points, as it would help us better address your concerns.

W3&Q3: Late-arriving Updates. We thank the reviewer for raising this issue. We propose a buffer-free model reconstruction framework for asynchronous federated learning that constructs reconstructed models for straggler models. Once a reconstructed model for a specific training task has been included in aggregation, the actual local model subsequently returned for that same task is excluded from future global aggregation. This prevents the actual model from being aggregated in addition to the reconstructed contributions for the same task.

If a client remains a straggler, its incomplete training task could still be represented in subsequent rounds by newly reconstructed models assembled from the latest received client models. This buffer-free design allows aggregation to proceed without waiting for straggling clients and avoids retaining the excluded late-arriving models in a queue for future aggregation.

W4&Q4: Theoretical analysis. We thank the reviewer for this suggestion. In future work, we will explore how reconstruction error affects global update bias and convergence in asynchronous aggregation.

Thank you for your detailed response and additional clarifications.

Regarding W1, I remain somewhat skeptical about the effectiveness of reconstructing pseudo models for straggling clients by assembling neural blocks from other clients. Although the proposed selection and fine-tuning procedures introduce additional adaptation, it is still unclear whether these reconstructed models can adequately represent the actual contributions of straggling clients. I believe this concern has not been fully addressed.

Regarding W2, I appreciate the additional visualizations and explanations. I think the importance of fingerprint stability may depend on the specific task and experimental setting. Therefore, I consider this a relatively minor concern.

Regarding W3, I have a more significant concern about directly discarding late-arriving genuine updates. In practical heterogeneous federated learning systems, some resource-constrained clients may consistently return their updates later than others. Under the proposed strategy, if their pseudo models have already been incorporated into aggregation, all their genuine updates could be discarded. Consequently, the information learned from these clients' local data may never directly contribute to the global model. Since the reconstructed pseudo models are assembled from other clients' models, they may not adequately preserve the unique information of these persistently slow clients. I believe this is a potentially serious limitation of the proposed approach, particularly under substantial system and data heterogeneity.

AuthorsOpenReview ↗

Indirect Straggler Contributions and Update Timeliness & Client-Specific Guidance for Model Reconstruction.

W3: Indirect Straggler Contributions & Update Timeliness.

Thank you for your insightful observation. First, stragglers can contribute to the global model through an indirect yet meaningful pathway. Specifically, our method leverages client fingerprints extracted from genuine local data distributions to guide similarity-based candidate selection during model reconstruction. Through this mechanism, even persistently slow clients can influence the reconstructed models and subsequent global aggregation via their historical fingerprints. Second, the value of genuine updates depends on both their knowledge and timeliness. Although late-arriving updates may contain unique client knowledge, they also be derived from outdated global model states. In particular, updates from persistently slow clients can introduce stale optimization directions, may hindering the current global optimization process.

In contrast, our method reconstructs pseudo-local models using timely local information, while leveraging fingerprint-based similarity constraints to guide candidate selection. This combination enables the global model to incorporate estimated straggler contributions in a timely manner while favoring information aligned with the current global optimization state.

W1: Client-Specific Guidance for Model Reconstruction.

We further clarify two key aspects of our method: how target-client fingerprints guide candidate selection and the conditions under which such guidance remains effective. First, our matching mechanism captures the sample relationships learned by the target client. Specifically, CKA measures the similarity between the pairwise sample similarity structures induced by two models. Thus, historical fingerprints characterize the representation structure learned by the target client through local training. Although candidate models are assembled from neural blocks of other clients, their selection is guided by the target-client fingerprint, enabling reconstruction to reflect client-specific characteristics. Second, effective candidate selection relies on the ability of historical fingerprints to preserve informative candidate rankings. Clients are warmed up on their respective local data, resulting in distinct feature responses. Consequently, even with an identical candidate pool, different client fingerprints can yield different candidate rankings. Under data heterogeneity, these fingerprints provide client-specific representation references learned from local data, allowing candidate selection to reflect the characteristics of each target client. Furthermore, the results in Appendix E demonstrate sufficient temporal consistency of fingerprints during subsequent training, allowing historical matching scores to remain informative about the relative similarity between candidate models and the current target-client model. This consistency supports the effectiveness of historical fingerprints in guiding candidate selection.

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

The paper studies asynchronous FL with heterogeneous client latency. Instead of discarding or down-weighting late updates, the authors build a pseudo local model on the server for each straggler and average it with the received models using equal weights. The pseudo model is assembled from blocks of the received models, which are grouped by CKA similarity and stitched together with an ordered borrowing rule. The candidates are fine-tuned on a held-out proxy set, and each straggler is assigned the candidate closest to its fingerprint, an activation profile recorded once before asynchronous training. On SVHN, CIFAR-10 and CIFAR-100 with ResNet-18, the method achieves the best mean accuracy in 24 of 27 settings against four AFL baselines.

Critical strengths

The main strength is the angle the paper takes on stragglers. Most AFL methods change how the received updates are weighted or corrected, while this work builds a stand-in for clients whose updates have not arrived yet, reusing blocks of models the server already holds. I have not seen model reassembly used this way before. If the stand-ins can be shown to carry real information about the stragglers, this would be a direction that complements existing AFL methods, and it is the main reason I see potential in the paper.

The ablation on matching also supports the design. Replacing the fingerprint-based selection with random assignment lowers accuracy in every tested setting, so how the candidates are selected matters and the gain does not come only from adding fine-tuned candidates to the average. This is the most direct evidence in the paper that the matching step contributes something, although it does not yet show that the selection is specific to each straggler.

Critical weaknesses
  1. No source code is provided, the paper does not say whether the three runs vary random seeds or data partitions, and many margins fall within one standard deviation.
  2. Efficiency is never measured. Each step runs pairwise CKA and fine-tunes several candidates per straggler, yet no server cost or wall clock time is reported, although the abstract lists efficiency as a goal.
  3. The baselines are few and mostly before 2025.
  4. Figure 1 is too small to read.
  5. The abstract claims that the pseudo models closely approximate the true local models received later, but the paper contains no such analysis, not even a comparison with simple stand-ins like the stale model or the current global model.
  6. Only the proposed method uses the IID proxy set and the extra server compute to fine-tune on it, which alone can ease non-IID drift. Without baselines given the same data and compute, the gain over the baselines cannot be attributed to straggler reconstruction.
  7. The experiments are not comprehensive. Only small image datasets are used, and the studies on client population, concurrency and backbone report the proposed method alone. Key factors are never varied, such as the proxy set size and distribution, the buffer size, and the number of local epochs, which is not even reported.
Questions for authors
  1. Is there any evidence that the pseudo models closely approximate the true local models received later, as claimed in the abstract? If not, could you report how close they are, compared with simply using the client's stale model or the current global model? Does this agreement (between pseudo and true model) change as training progresses while the fingerprints stay fixed?
  2. How are the pending clients in S_t selected? Can the same pending task be represented by a substitute in several aggregation steps, and how is its real model handled when it arrives?
  3. Do the three runs use different random seeds, different data partitions, or both? Will the code be released?
  4. How many local epochs, what batch size and which optimizer are used for client training?
  5. How do the baselines perform when given the same proxy set and server compute, for example by fine-tuning the global model on the proxy set at each step?
  6. What are the server computation and wall clock time per aggregation step, and how does accuracy compare with the baselines when plotted against wall clock time? Are the costs of warm-up, block search and candidate fine-tuning counted in the simulated time?
  7. How sensitive are the results to the size and distribution of the proxy set, the buffer size and the number of local epochs? Do the baselines show the same trends in the studies on client population and concurrency?
5 comments
AuthorsOpenReview ↗

We clarify the experimental setup and provide additional evaluations of computational overhead, reconstruction reliability, and public-data usage, alongside comparisons with FedASRA. We also extend experiments to larger datasets, client populations, and concurrency levels, analyze parameter sensitivity.

W1&Q3&Q4: Experimental Setup. We run each experiment three times with different training random seeds while keeping the client data partition fixed. Client training uses 20 local epochs, a batch size of 32, andSGD. We will release the source code later.

W2&Q6: Computational Overhead. To further assess the computational efficiency of our method, we report the total wall-clock time required for each method to converge, defining convergence as accuracy stabilization with no further improvement over 10 consecutive rounds. Appendix.F reports the time to convergence for each method, while Appendix.F details the costs of the three stages : client warm-up, block search, and candidate fine-tuning. These results demonstrate competitive overall time to convergence under the evaluated setting. This advantage stems from the timely reconstruction of straggler models, which facilitates global convergence.

W3: Comparison with 2026 Baseline. To evaluate our method against recent advances, we include FedASRA as an additional baseline. FedASRA improves client selection and corrects stale gradients within a conventional asynchronous framework. Appendix.G presents the comparison results, which further support the effectiveness of our model reconstruction method. [1]Z. Liu, M. Li, B. Zhang, X. Liu and R. Zhu, "Taming Staleness in Asynchronous Federated Learning With System Heterogeneity," in IEEE Transactions on Parallel and Distributed Systems, vol. 37, no. 10, pp. 2268-2282, Oct. 2026, doi: 10.1109/TPDS.2026.3725721.

W4: Figure Readability.. We will increase the font size in Fig.1 to improve readability.

W5&Q1: Reconstruction Reliability. We thank the reviewer for this concern. We evaluate reconstruction reliability by pairing each reconstructed model generated at aggregation time with the actual local model subsequently returned by the straggling client. We assess the similarity between reconstructed and actual local models in terms of both representations and model updates. Representational similarity reflects agreement in their feature structures on the same proxy inputs, while update cosine similarity reflects the alignment of their optimization directions in parameter space. We feed the same proxy samples into both models, extract their feature representations, and compute CKA similarity. We also compute each model update and measure the cosine similarity. Appendix.H indicate that our reconstructed models exhibit similarity to the actual local models in both feature representations and parameter-update directions. This similarity stems from the integration of our search algorithm and matching mechanism. The ordered block-borrowing search generates diverse candidate combinations, while similarity matching against cached fingerprints further guides the selection of suitable reconstructed models for straggling clients.

W6&Q5: Public datasets. To ensure that our method and the baselines use the same overall training data, we conduct two control experiments to examine the effect of public-data usage (Appendix.I). First, we restore public data to baselines. Our method achieves the highest accuracy in most configurations. Second, we fine-tune each baseline’s global model on the same public dataset after every aggregation step. Our method maintains the highest accuracy in most configs. These results indicate that our performance gains cannot be attributed solely to the use of public data and further support the contribution of our model reconstruction method.

W7&Q7: Extended Evaluation and Sensitivity. To evaluate our method on larger datasets, we extend our experiments to Tiny ImageNet, CelebA and FEMNIST. Dataset descriptions and details of the non-IID partitions are provided in Appendix J.1. Our method achieves competitive accuracy in most configurations. These results indicate that our method supports model reconstruction on larger datasets. To examine baseline performance with larger client populations and different concurrency levels, we provide additional comparisons in Appendix J.2. Overall, baseline accuracy declines in most configurations as concurrency increases, although the magnitude and direction of change vary across methods and datasets. To examine the sensitivity of our method to proxy-data size and distribution, as well as the number of local epochs, we conduct additional experiments reported in Appendix J.3. The observed trends may reflect the roles of proxy data and local training in model reconstruction. Sufficient local training helps clients learn informative features, providing better-trained network blocks for reconstruction. Our method is buffer-free and therefore has no aggregation buffer size to tune.

Q2: Straggler Selection and Continuous Model Reconstruction. Further explanation can be found in Reply below (also in Appendix K).

AuthorsOpenReview ↗

For easy viewing of the results, experimental results in Appendix F,G,H,I.

Appendix.F:

Table 5. Time to convergence.

MethodTotal time (s)
FedBuff1613.107
CA2FL1664.173
FADAS1523.129
AdamasFL1161.48
Ours1256.08

Table 6. Component costs for our method.

Componenttime (s)
Client warm-up285.46
Block clustering and search527.55
Candidate fine-tuning374.77

Appendix.G:

Table 7. Accuracy comparison with FedASRA (%). R denotes Real-world (Stragglers 10–30%); U30 denotes Uniform w/ Delay (Stragglers 30–40%); and U40 denotes Uniform w/ Delay (Stragglers 40–60%).

SettingMethodSVHN α=0.6SVHN α=0.3SVHN α=0.1CIFAR-10 α=0.6CIFAR-10 α=0.3CIFAR-10 α=0.1CIFAR-100 α=0.6CIFAR-100 α=0.3CIFAR-100 α=0.1
RFedASRA92.5392.1387.3878.6475.2772.4860.2951.2852.83
ROurs95.2594.5592.5687.9686.4074.5668.9564.6263.37
U30FedASRA91.2390.9185.9183.1680.2570.0457.8053.3838.32
U30Ours95.3494.4493.7685.5983.1072.3364.9462.7758.97
U40FedASRA91.2390.0985.0178.1475.7370.9760.9148.5249.10
U40Ours95.4194.1293.7484.4982.2672.5162.7661.0754.30

Appendix.H:

Table 8. Reconstruction similarity.

RoundsCKACosinePositive cosine (%)
1–1000.56560.75895.49
101–2000.56500.64797.88
201–3000.57230.54999.66
301–4000.55480.45498.19
401–5000.53490.36899.63

Appendix.I:

Table 9. Accuracy with baseline (%).

Dataset (α)FedBuffCA2FLFADASAdamasFLOurs
SVHN (0.6)93.68 / 93.24 / 92.9792.21 / 91.80 / 93.5692.63 / 92.38 / 89.2090.35 / 90.50 / 89.7895.25 / 95.34 / 95.41
SVHN (0.3)92.63 / 91.76 / 94.6493.45 / 94.09 / 71.6093.29 / 89.62 / 90.6994.01 / 90.27 / 90.2394.55 / 94.44 / 94.12
SVHN (0.1)89.39 / 88.32 / 91.0191.23 / 91.08 / 12.9091.49 / 86.50 / 79.7491.82 / 89.05 / 89.6592.56 / 93.76 / 93.74
CIFAR-10 (0.6)84.05 / 82.34 / 80.1784.35 / 59.89 / 54.6183.51 / 85.56 / 82.9162.42 / 62.83 / 62.1987.96 / 85.59 / 84.49
CIFAR-10 (0.3)79.92 / 75.66 / 72.4482.55 / 47.50 / 46.0281.76 / 82.85 / 82.8761.22 / 61.34 / 60.5286.40 / 83.10 / 82.26
CIFAR-10 (0.1)69.47 / 55.46 / 69.2972.57 / 68.24 / 39.5173.59 / 72.79 / 75.0044.52 / 56.19 / 56.9174.56 / 72.33 / 72.51
CIFAR-100 (0.6)62.27 / 63.97 / 63.6367.03 / 38.58 / 36.1459.38 / 60.05 / 59.5223.06 / 22.67 / 22.6968.95 / 64.94 / 62.76
CIFAR-100 (0.3)60.84 / 61.05 / 61.1364.22 / 32.14 / 33.5754.80 / 58.28 / 57.2922.11 / 22.73 / 22.4064.62 / 62.77 / 61.07
CIFAR-100 (0.1)55.51 / 37.76 / 37.8863.78 / 47.87 / 23.2751.30 / 22.02 / 22.4622.33 / 21.02 / 19.6863.37 / 58.97 / 54.30

Table 10. Accuracy with proxy-data fine-tuning of baseline global models (%).

DatasetαFedBuffCA2FLFADASAdamasFLOurs
SVHN0.694.34 / 93.47 / 94.2894.47 / 94.42 / 94.4892.74 / 92.40 / 92.4587.53 / 87.77 / 87.6695.25 / 95.34 / 95.41
SVHN0.392.73 / 93.51 / 93.2893.13 / 94.26 / 94.2992.65 / 92.08 / 92.3187.53 / 87.52 / 87.5294.55 / 94.44 / 94.12
SVHN0.192.67 / 93.17 / 93.5193.68 / 93.97 / 93.9891.58 / 91.04 / 90.6287.32 / 86.85 / 87.4392.56 / 93.76 / 93.74
CIFAR-100.684.33 / 85.10 / 84.8786.22 / 86.65 / 83.4882.32 / 77.99 / 76.8056.05 / 57.04 / 57.3187.96 / 85.59 / 84.49
CIFAR-100.381.55 / 82.52 / 82.4986.79 / 82.95 / 81.4780.11 / 75.03 / 74.8656.51 / 56.28 / 57.2686.40 / 83.10 / 82.26
CIFAR-100.171.65 / 72.36 / 78.6872.52 / 71.18 / 79.5373.26 / 68.17 / 66.8153.76 / 54.65 / 53.2674.56 / 72.33 / 72.51
CIFAR-1000.659.82 / 60.15 / 60.0765.11 / 64.62 / 61.1152.85 / 44.28 / 40.1519.73 / 19.59 / 21.1168.95 / 64.94 / 62.76
CIFAR-1000.358.07 / 58.39 / 57.4064.07 / 61.45 / 57.8550.94 / 43.74 / 39.9118.54 / 17.65 / 19.9164.62 / 62.77 / 61.07
CIFAR-1000.152.95 / 53.19 / 51.1162.42 / 59.73 / 53.1849.09 / 40.93 / 38.8119.87 / 16.39 / 16.4163.37 / 58.97 / 54.30
AuthorsOpenReview ↗

For easy viewing of the results, experimental results in Appendix J.1,J.2,J.3,andK

Table 12. Accuracy on larger datasets (%).

MethodTiny ImageNet α=0.6Tiny ImageNet α=0.3Tiny ImageNet α=0.1CelebAFEMNIST
FedBuff28.5626.1620.8789.6985.98
CA2FL14.8113.6714.8177.9738.34
FADAS24.3420.2914.2071.8284.32
AdamasFL10.4010.5510.4083.3983.97
Ours29.6828.3820.6890.6587.75

Table 13. Accuracy with 500 clients and 10% participation per round (%).

MethodSVHN α=0.6SVHN α=0.3SVHN α=0.1CIFAR-10 α=0.6CIFAR-10 α=0.3CIFAR-10 α=0.1CIFAR-100 α=0.6CIFAR-100 α=0.3CIFAR-100 α=0.1
FedBuff93.6393.4390.0384.1882.6074.0953.1652.7551.81
CA2FL94.2393.0890.6283.8474.1466.8350.8450.3552.96
FADAS92.4192.0391.1581.9079.6074.1148.8148.0646.96
AdamasFL90.3193.4890.9663.2961.3853.5122.5922.1721.91
Ours93.7593.6791.4784.2283.6572.2162.5260.9155.12

Table 14. Accuracy at different concurrency levels (%).

ConcurrencyMethodSVHN α=0.6SVHN α=0.1CIFAR-10 α=0.6CIFAR-10 α=0.1CIFAR-100 α=0.6CIFAR-100 α=0.1
15FedBuff93.2091.9374.2378.3745.5843.62
15CA2FL91.2284.4670.0865.1343.7741.20
15FADAS92.2586.7077.2288.9243.1941.30
15AdamasFL89.5087.0859.8887.6821.2521.01
15Ours95.7593.7489.7979.3867.7055.01
20FedBuff93.4891.3778.5163.4042.5741.65
20CA2FL83.3165.1377.2966.6243.3535.60
20FADAS90.6190.2273.1664.6540.0538.77
20AdamasFL89.6887.6861.1142.3721.6219.90
20Ours94.5793.7192.5078.8368.1665.55

Table 15. Sensitivity to proxy data and local training (%).

SettingSVHN α=0.6SVHN α=0.3SVHN α=0.1CIFAR-10 α=0.6CIFAR-10 α=0.3CIFAR-10 α=0.1
Default: 10%, IID, 20 epochs95.2594.5592.5687.9686.4074.56
Proxy size: 2.5%94.0693.9592.0084.5881.6472.52
Proxy size: 5%94.2194.0892.8884.1385.3273.20
Proxy distribution: non-IID94.0594.0791.8784.5283.4872.96
Local epochs: 592.9590.2286.1483.0682.8570.86
Local epochs: 1090.0689.6687.7685.0483.7172.60
Local epochs: 1590.4789.2890.2285.8685.6573.04

Q2: Straggler Selection and Continuous Model Reconstruction.

In each of the three simulation scenarios, the straggler set (S_t) is randomly sampled at each round according to the corresponding scenario settings. If a client remains a straggler, the same task can be represented by newly reconstructed models in multiple aggregation rounds. At each step, we construct a new model for that straggling client by reassembling blocks from the latest received client models. This allows reconstruction to incorporate newly available information across rounds. Reconstruction stops when the client is no longer classified as a straggler. Once a training task has been represented by a reconstructed model, its straggler model is excluded from global aggregation. This prevents the genuine model from being aggregated in addition to the reconstructed contributions for the same task. Our buffer-free design allows aggregation to proceed without waiting for straggling clients and avoids retaining these excluded late-arriving models in a queue for future aggregation.

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

This paper proposes a bi-granularity aggregation framework for asynchronous federated learning, aiming to incorporate information from straggling clients without waiting for their updates. It reconstructs pseudo local models by reassembling neural blocks from available client models and selecting candidates through cached activation fingerprints. The reconstructed models are aggregated with received local models. Experiments on three datasets demonstrate improved accuracy under different data heterogeneity and asynchronous settings.

Critical strengths
  1. The paper addresses an interesting problem of underutilized straggler information in asynchronous federated learning.

  2. The idea of reconstructing unavailable client models through neural-block reassembly is interesting.

  3. The experimental coverage is relatively comprehensive, including state-of-the-art baselines, different datasets, straggler settings, client populations, and ablation studies.

Critical weaknesses
  1. The main concern is novelty. The neural-block clustering, ordered model reassembly, and candidate matching are closely related to existing model reassembly methods such as pFedClub. The main contribution appears to be adapting these techniques to asynchronous FL, while the fundamental technical differences remain insufficiently justified.

  2. The proposed bi-granularity weighted aggregation is not clearly formulated. Eq. (2) simply averages received and reconstructed models with uniform weights. There is no explicit weighting mechanism at either the model or neural-block level, which makes the claimed aggregation design questionable.

  3. The reliability of pseudo-model reconstruction remains a concern. Matching candidates with static activation fingerprints does not necessarily guarantee similarity to the actual pending local models, especially as training progresses. Direct quantitative reconstruction-error analysis and corresponding theoretical justification are lacking.

  4. The computational overhead is substantial. The method requires client warm-up and repeatedly generates and fine-tunes numerous candidate models on the server. However, the experiments mainly report accuracy against communication rounds without evaluating the actual training-time overhead, which raises concerns about its practicality for asynchronous FL.

  5. The experimental improvements may partly benefit from additional proxy-data training. Although the same data are excluded from baseline clients, the proposed method uses them for extensive server-side calibration. Comparisons with simpler model substitution strategies under matched data and computational budgets are expected.

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

We clarify our asynchronous design and aggregation weights and provide additional evaluations of reconstruction reliability, computational overhead, and the impact of proxy data.

W1: Asynchronous Design. We design a model reconstruction method for asynchronous federated learning to approximately compensate for the updates from stragglers in the current aggregation round. As training progresses, differences in client response times increase the discrepancies between the global model obtained through asynchronous aggregation and the one obtained under full participation. To address this, we construct candidate models by reassembling blocks from limited received models. Then we use the local fingerprints cached during warm-up to identify a matching reconstructed model for each straggling client. With this design, we aim to bring the resulting global model closer to the one obtained under full participation.

W2: Aggregation Weights. We assign equal weight to each input model, including received client models and reconstructed models. Uniform averaging is the foundational implementation adopted in this work, rather than a necessary constraint of the bi-granularity weighting framework. The formulation in Eq. (2) can be generalized to other weighting rules that adjust the respective contributions of received and reconstructed models.

W3: Reconstruction Reliability. We thank the reviewer for this concern. We evaluate reconstruction reliability by pairing each reconstructed model generated at aggregation time with the actual local model subsequently returned by the straggling client. We assess both representations and model updates. We feed the same proxy samples into both models, extract their feature representations, and compute CKA similarity. We also compute each model update and measure the cosine similarity. The results in Table 1 indicate that our reconstructed models exhibit a degree of similarity to the actual local models in both feature representations and parameter-update directions.

Table 1. Reconstruction similarity.

RoundsCKACosinePositive cosine (%)
1–1000.56560.75895.49
101–2000.56500.64797.88
201–3000.57230.54999.66
301–4000.55480.45498.19
401–5000.53490.36899.63

W4: Computational overhead. We thank the reviewer for this concern. We report the total wall-clock time required for each method to converge. We define convergence as accuracy stabilization with no further improvement over 10 rounds. Table 2 reports our results. These results show that our approach achieves competitive time to convergence, supporting its practical feasibility under the evaluated setting. This advantage stems from the timely reconstruction of straggler models, which facilitates global convergence.

Table 2. Time to convergence.

MethodTotal time (s)
FedBuff1613.107
CA2FL1664.173
FADAS1523.129
AdamasFL1161.48
Ours1256.08

W5: Additional Public Dataset. To examine this factor, we give the 10% proxy data to the baselines’ training sets, giving all methods access to the same total amount of training data. As shown in Table 3, our method achieves the highest accuracy in most settings. These results further support the effectiveness of our method and indicate that its performance gains cannot be attributed solely to the use of additional data.

Table 3. Accuracy with Additional Public Dataset(%). R / U30 / U40: R = Real-world (10–30% stragglers); U30 = Uniform w/ Delay (30–40% stragglers); U40 = Uniform w/ Delay (40–60% stragglers).

Dataset (α)FedBuffCA2FLFADASAdamasFLOurs
SVHN (0.6)93.68 / 93.24 / 92.9792.21 / 91.80 / 93.5692.63 / 92.38 / 89.2090.35 / 90.50 / 89.7895.25 / 95.34 / 95.41
SVHN (0.3)92.63 / 91.76 / 94.6493.45 / 94.09 / 71.6093.29 / 89.62 / 90.6994.01 / 90.27 / 90.2394.55 / 94.44 / 94.12
SVHN (0.1)89.39 / 88.32 / 91.0191.23 / 91.08 / 12.9091.49 / 86.50 / 79.7491.82 / 89.05 / 89.6592.56 / 93.76 / 93.74
CIFAR-10 (0.6)84.05 / 82.34 / 80.1784.35 / 59.89 / 54.6183.51 / 85.56 / 82.9162.42 / 62.83 / 62.1987.96 / 85.59 / 84.49
CIFAR-10 (0.3)79.92 / 75.66 / 72.4482.55 / 47.50 / 46.0281.76 / 82.85 / 82.8761.22 / 61.34 / 60.5286.40 / 83.10 / 82.26
CIFAR-10 (0.1)69.47 / 55.46 / 69.2972.57 / 68.24 / 39.5173.59 / 72.79 / 75.0044.52 / 56.19 / 56.9174.56 / 72.33 / 72.51
CIFAR-100 (0.6)62.27 / 63.97 / 63.6367.03 / 38.58 / 36.1459.38 / 60.05 / 59.5223.06 / 22.67 / 22.6968.95 / 64.94 / 62.76
CIFAR-100 (0.3)60.84 / 61.05 / 61.1364.22 / 32.14 / 33.5754.80 / 58.28 / 57.2922.11 / 22.73 / 22.4064.62 / 62.77 / 61.07
CIFAR-100 (0.1)55.51 / 37.76 / 37.8863.78 / 47.87 / 23.2751.30 / 22.02 / 22.4622.33 / 21.02 / 19.6863.37 / 58.97 / 54.30

Thank the authors for the detailed responses. The additional experiments on reconstruction reliability, computation time, and proxy data are helpful. I'd like to strongly encourage such discovery in this interesting direction. However, my main concerns about the weighting mechanism remain. The declining cosine similarity between reconstructed and actual updates also raises further concerns about reconstruction reliability. I really hope the authors could work further to improve the work for future acceptance. Overall, I think the work is worth being published. I will maintain my rating at 3.

AuthorsOpenReview ↗

Thank you once again for your encouragement, positive assessment, and constructive feedback! In future work, we will investigate the underlying causes of these two issues and explore strategies to address them.

Thank you once again for your encouragement, positive assessment, and constructive feedback! In future work, we will investigate the underlying causes of these two issues and explore strategies to address them.

Comments on the paper

2 comments
AuthorsOpenReview ↗

To All Reviewers: Updated Appendix

Thanks to all reviewers for their constructive feedback! Main appendix updates and clarifications:

  1. Reviewer gmPY. Clarification of our asynchronous design and bi-bi-granularity weight aggregation; reconstruction model reliability; computational cost; fair use of public datasets for the baselines.
  2. Reviewer qUqe. Experimental setup; computational cost (Appendix F, Table5,6); additional comparisons with 2026 methods (Appendix G, Table7); reconstruction model reliability (Appendix H, Table8); fair use of public datasets for baselines in two ways, including public datasets for clients and global model (Appendix I, Table9,10); larger datasets performance(Appendix J.1, Table12), large-scale clients and concurrency performance for baselines; public dataset size, distribution, local epochs sensitivity experiments for our method (Appendix J.2–J.3, Table13,14,15); continuous model reconstruction for Stragglers (Appendix K).
  3. Reviewer ZqK6. Distinction between model reassembly and block-wise reweighting; warm-up fingerprint stability visualizations (Appendix E, Figure 5); late-update handling.
AuthorsOpenReview ↗

Rebuttal Summary and Request for Feedback

Dear ACs and Reviewers,

Thank you for your time and constructive feedback. As the discussion period draws to a close, we would like to summarize our key clarifications and additional experimental results, and welcome any remaining feedback.

In response to the reviewers' comments, we have provided the following clarifications and additional evidence:

  • Reviewer gmPY: The reviewer raised questions regarding the asynchronous design, bi-granularity weighting, reconstruction reliability, computational overhead, and fairness in the use of public data. We clarified that our method approximately compensates for straggling clients' contributions to mitigate the growing discrepancy between asynchronous aggregation and aggregation with more complete client participation. We also explained that equal-weight averaging is the simple implementation adopted in this work, while our framework allows other weighting strategies. Additional experiments evaluate the similarity between reconstructed and actual straggler models in terms of feature representations and parameter-update directions, and compare computational overhead. Furthermore, we provide the same public data to all baselines to ensure fair comparisons under equal overall data budgets.

  • Reviewer qUqe: The reviewer requested clearer experimental details and broader evaluations of recent baselines, reconstruction reliability, efficiency, public-data fairness, scalability, and sensitivity, as well as clarification of straggler selection. We specified the experimental hyperparameters and added comparisons with a 2026 baseline. In addition to the reconstruction similarity evaluation and computational cost comparison, we conducted two additional public-data experiments: incorporating the same public data into baseline client training and fine-tuning baseline global models. We extended our evaluation to larger datasets, including Tiny ImageNet, CelebA, and FEMNIST, and evaluated the baselines with larger client populations and higher concurrency levels. We also conducted sensitivity analyses of public-data size, public-data distribution, and the number of local epochs. Finally, we clarified that stragglers are randomly sampled in each round, rather than predetermined.

  • Reviewer ZqK6: The reviewer questioned the distinction from block-wise reweighting, fingerprint stability, the handling of late-arriving updates, and theoretical justification. We clarified that our reconstruction pipeline differs from direct block-wise reweighting by considering whole-model behavior and parameter adaptation. We added warm-up fingerprint visualizations showing similar spatial response patterns in highlighted channels despite variations in activation intensity and local details, providing qualitative evidence of fingerprint stability. We further clarified that once a training task has been represented in aggregation through reconstructed models, its corresponding late-arriving update is excluded from subsequent aggregation. This buffer-free design enables aggregation without waiting for stragglers or maintaining a queue of excluded late-arriving updates. We leave a rigorous theoretical analysis for future work.

We would also like to reiterate the three central contributions of our submission.

  • Bi-granularity weighted aggregation. Our framework combines model-level aggregation of received client models with neural-block-level reconstruction of pseudo-local models, enabling asynchronous federated learning to timely incorporate estimated contributions from straggling clients.

  • Model reconstruction from limited available models. Our ordered block-borrowing search explores diverse neural-block combinations, while cached client fingerprints guide candidate selection, enabling model reconstruction without access to the corresponding straggler models.

  • Extensive empirical evaluation. Our method achieves competitive accuracy and convergence across diverse tasks. Additional experiments further demonstrate its effectiveness on larger datasets, with larger client populations, under higher concurrency levels settings.

Thank you again for the thoughtful discussion. We sincerely appreciate your consideration of these clarifications and additional results in your final assessments, and welcome any further questions before the discussion closes.

Sincerely,

The Authors of Submission 31446

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.