acceptodds
Under review as a conference paper at ICLR 2027

Aggregation-Faithful Value Adaptation for Federated Vision-Language Models

Abstract

Federated adaptation of large vision-language models (VLMs) is appealing for privacy-sensitive applications but must contend with client data heterogeneity, strict communication budgets, and on-device inference constraints. Recent methods predominantly adopt lightweight prompt or adapter interfaces, yet averaging their parameters may fail to preserve the functional changes learned across heterogeneous clients. We therefore ask: can robust federated VLM adaptation be achieved through an aggregation-faithful intrinsic interface? We answer this question with TriVA (Tri-State Value Adaptation), a module-free framework that fine-tunes only native value biases in both visual and textual encoders. TriVA operates on existing backbone parameters and introduces no additional inference-time modules. Our analysis shows that this intrinsic interface provides substantially higher aggregation fidelity than the tested prompt- and adapter-based alternatives. To make this aggregation-friendly interface effective under heterogeneous clients without excessive synchronization, we further introduce a one-shot calibration mechanism that assigns value-bias groups to three persistent states: Frozen groups preserve pretrained values, Shared groups participate in federated aggregation, and Private groups retain client-specific adaptation locally. The resulting assignments remain fixed throughout training, with only shared groups exchanged for aggregation. Extensive experiments across heterogeneous federated benchmarks demonstrate strong adaptation performance with a compact trainable parameter set and substantially reduced communication cost. Semantic segmentation experiments further demonstrate the applicability of TriVA beyond image classification to dense prediction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Federated Learning, Vision-Language Models, Personalized Federated Learning, Parameter-Efficient Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ET3aOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TriVA (Tri-State Value Adaptation) for parameter-efficient federated adaptation of CLIP-style vision-language models. TriVA updates only native value-projection biases in the visual and textual encoders. Its motivation is that, for fixed attention-layer inputs, value-bias updates introduce additive corrections that are preserved exactly under parameter averaging. To accommodate heterogeneous clients, a one-shot calibration procedure assigns attention-head bias groups to three persistent states: Frozen, Shared, or Private. Assignments depend on the estimated utility of local adaptation and the loss incurred by replacing local parameters with their globally averaged counterparts. Only Shared groups participate in subsequent communication. Experiments cover few-shot classification, domain heterogeneity, larger client populations, and a CamVid segmentation extension. The paper reports competitive accuracy with a small adaptation interface and low steady-state communication costs.

Critical strengths

(1) The paper correctly separates whether parameter averaging reproduces functional averaging from whether sharing is beneficial for heterogeneous clients. This distinction provides a coherent motivation for combining the adaptation interface with selective synchronization.

(2) Equation (1) establishes an exact additive property for fixed layer inputs under the stated attention assumptions. The manuscript appropriately acknowledges that this property does not imply end-to-end linearity when adapting multiple layers and encoders.

(3) Updating existing value biases avoids inserting prompts or adapters. This offers a practical deployment advantage, particularly when model architecture and inference implementation should remain unchanged.

(4) The Frozen/Shared/Private assignments distinguish adaptation utility from synchronization suitability. Evaluating replacements on local probes is more interpretable than assigning roles solely through parameter magnitude or predetermined layer partitions.

(5) Experiments include label-distribution heterogeneity, domain shifts, different supervision budgets, and 100-client settings. The paper also acknowledges settings where TriVA does not achieve the best accuracy, including CIFAR-100 and some synchronization ablations.

Critical weaknesses

(1) Figure 1 provides precise merge-error values for prompts, adapters, and value biases, but the manuscript does not adequately specify the diagnostic dataset, client configurations, optimization settings, adaptation budgets, or variability underlying these measurements. The normalized metric also does not fully eliminate differences in adaptation magnitude. Under a locally smooth parameter-to-function mapping, smaller updates can yield smaller normalized nonlinear aggregation errors. Comparisons should therefore control for functional displacement or local adaptation quality, rather than relying on the reported normalized scores alone.

(2) On Stanford Cars, removing Private parameters improves accuracy from 79.34% to 79.50%, while removing Shared parameters yields 79.07% with no steady-state communication. The complete method offers a possible accuracy–communication compromise, but these results do not establish that all three states are consistently necessary. Broader comparisons against local-only value-bias tuning, binary Frozen/Private and Shared/Private assignments, random assignments with matched state proportions, and budget-matched alternatives are needed.

(3) Roles are determined using three probe batches after initial calibration and remain fixed throughout training. Their reliability is especially uncertain under very limited local supervision. The paper does not report sensitivity to probe sampling, calibration duration, the active-ratio cap, or the stability of role assignments across seeds. There is also a mismatch between calibration and subsequent aggregation: calibration evaluates replacement by an average over all clients, whereas later updates average only clients that mark a group Shared. The initial synchronization-cost estimate therefore evaluates a different composition from the one used during training.

(4) Proposition 3.1 follows from the Lipschitz continuity of cross-entropy and a decomposition involving the functional-average risk. It clarifies two sources of synchronization error, but does not establish that value-bias adaptation minimizes either term or that TriVA improves population risk. The joint-assignment analysis bounds cross-group interactions through mixed Hessian terms, but these quantities are not estimated empirically. Consequently, the sufficient condition does not establish that the actual assignments are safe, particularly when useful groups are frozen by the active-ratio cap.

(5) Important details remain unclear, including learning rates, optimizer settings, client participation, probe construction, repeated runs, and the CamVid adaptation pipeline and partitioning protocol. These omissions matter because several reported margins are small: on DomainNet-100, TriVA exceeds pFedMMA by only 0.06 % . Communication figures explicitly measure steady-state parameter payloads, which is reasonable, but they exclude the initial all-group calibration exchange and other setup costs. Calibration also requires repeated counterfactual evaluations across bias groups. Total communication, calibration runtime, and training memory should be reported alongside steady-state costs.

Questions for authors

(1) What exact protocol produces the aggregation-fidelity results in Figure 1? Please report diagnostic data, optimization settings, update magnitudes, local performance, and variability.

(2) Does the fidelity advantage persist when interfaces are matched by functional displacement, local accuracy, or trainable parameter count?

(3) How does TriVA compare with local-only value-bias adaptation and random or binary role assignments across all benchmarks?

(4) How stable are the assignments across probe samples and random seeds? How sensitive are results to calibration duration and the active-ratio cap?

(5) Why should a synchronization probe using the all-client average predict the benefit of later aggregation over only Shared clients?

(6) Can the authors measure cross-group interaction effects and the actual joint loss change induced by calibration assignments?

(7) What are the complete communication and computational costs, including calibration? Please also specify the segmentation pipeline, client participation protocol, and independent-run statistics.

1 comment
AuthorsOpenReview ↗

Clarifications on Aggregation Fidelity, Tri-State Ablations, Cross-Group Interactions, and Calibration Reliability

W1/Q1/Q2: Aggregation fidelity and matched adaptation comparison

Figure 1 uses Stanford Cars with 10 clients, 16 shots per class, and 8,041 test images. All interfaces use SGD (LR 0.004, momentum 0.9, weight decay 0.0005), five local epochs, and data-weighted FedAvg. Each round uses the global model before local training as the reference. We report medians over 50 client-round measurements from rounds 46–50, with interquartile ranges (IQR).

InterfaceParams.Merge Error ↓IQR [Q1, Q3]Update L2DisplacementLocal Acc. (%)Local CE
Prompt8,1920.4123[0.4020, 0.4204]0.90123.287381.370.5771
Adapter140,4160.2085[0.2006, 0.2141]0.53430.930074.750.7622
Value-bias15,3600.0148[0.0143, 0.0151]0.70311.196881.060.5484

We further compare four interfaces under matched local CE improvement using 1,635 separate validation images.

InterfaceLocal CE GainMerge Error ↓
Prompt0.0479950.017337
Adapter0.0477920.005789
Output-bias0.0478460.005412
Value-bias0.0479000.003580

Value-bias achieves the lowest normalized merge error in both comparisons, supporting its aggregation-fidelity advantage under comparable local adaptation quality.

W2/Q3: Tri-state ablation comparisons

We extended the tri-state ablations to all five single-domain benchmarks, comparing TriVA with local-only adaptation, binary role assignments, and matched-budget random allocation.

MethodCaltech101UCF101Stanford CarsSUN397Food101
TriVA (Ours)96.2682.7679.3474.7596.44
Local-Matched / Binary Frozen–Private96.2282.8479.0774.1496.42
Matched Random96.2583.0878.9474.3296.41
Binary Shared–Private96.1882.5478.7274.0996.39

TriVA achieves the highest accuracy on four of the five benchmarks and consistently outperforms both binary role assignments, supporting the effectiveness of counterfactual tri-state synchronization.

W3/Q4/Q5: Calibration reliability and synchronization consistency

We evaluated calibration sensitivity to probe size, sampling seeds, calibration duration, and the active-ratio cap on Stanford Cars.

FactorSettingsLast-10 Acc. (%)
Probe batches1 / 3 / 579.25 / 79.34 / 79.31
Probe seeds101 / 102 / 10379.30 / 79.39 / 79.32
Calibration epochs1 / 5 / 1079.31 / 79.34 / 78.89
Active-ratio cap0.50 / 0.75 / 1.0079.71 / 79.34 / 79.60

Role agreement ranges from 50.33% to 55.21% across probe sizes and from 50.6% to 56.3% across sampling seeds. Although individual assignments vary, accuracy remains relatively stable under the tested probe configurations.

To examine the difference between all-client calibration and subsequent Shared-only aggregation, we recomputed replacement costs using actual selectively aggregated parameters. At rounds 10, 25, and 50, 62.05%–68.41% of initially Shared groups retain non-positive replacement costs. Selected-client recalibration achieves 79.38% accuracy, compared with 79.34% for the original one-shot calibration.

One-shot calibration achieves comparable accuracy to selected-client recalibration without additional overhead.

W4/Q6: Joint assignment and cross-group interactions

Proposition 3.1 characterizes synchronization risk rather than guaranteeing non-increasing joint loss. We empirically evaluated cross-group interactions and actual joint calibration loss changes on SUN397.

MetricSUN397
Sum of individual effects−0.03768
Joint ΔCE+0.00853
Interaction residual+0.04621
Joint ΔCE after cap restoration+0.00531

For the five clients affected by the active-ratio cap, restoring 165 frozen groups with positive adaptation utility reduces their mean joint ΔCE from +0.00659 to just +0.00013.

This result supports the utility of TriVA's adaptation-utility estimates in identifying beneficial parameter groups.

W5/Q7: Reproducibility and total overhead

Cars: 10/10 clients participate; probes sample 96 training images/client without replacement. Profiling: 18.57 min calibration (15.45 min probing), 180.27 min total, 5.03 GiB peak GPU memory. Separate run: 10.83 MB bidirectional traffic (1.53 MB calibration; 9.30 MB later), including metadata.

CamVid: 367/101/233 train/val/test images, 11 classes, 10 clients (Dirichlet β=0.5 by dominant non-void class); CLIP ViT-B/16 + shared convolutional decoder, 224×224 inputs, SGD LR 0.004, 50 rounds × 5 epochs, batches 8/16, pixel CE ignoring void; last-10 mIoU.

Appendix Table 8 reports three-seed mean±SD for five single-domain datasets. Due to compute limits, DomainNet-100 uses one seed; three-seed results will be included in the final version.

Reviewer_axJKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TriVA, a personalized federated adaptation method that fine-tunes native value-projection biases in both encoders of a pretrained vision-language model. It motivates this interface through the additive effect of value-bias updates on attention outputs and evaluates the discrepancy between parameter averaging and averaging prediction changes. A one-time calibration assigns each attention-head bias group to Frozen, Shared, or Private using local loss-based counterfactual replacements. Only Shared groups participate in subsequent aggregation. Experiments with CLIP ViT-B/16 report competitive classification accuracy and reduced steady-state communication, with an additional semantic segmentation evaluation. The theoretical analysis distinguishes aggregation fidelity from the benefit or harm of replacing personalized predictions with a shared model.

Critical strengths
  1. A compact adaptation design with practical communication benefits. Updating existing value biases avoids additional inference modules, while retaining some groups locally provides a straightforward mechanism for personalization. The reported results on the five single-domain datasets, together with the low steady-state communication payload, make this a promising approach for resource-constrained federated adaptation.

  2. A useful distinction between aggregation fidelity and synchronization utility. The paper correctly separates agreement between parameter and functional averaging from whether a shared prediction function benefits heterogeneous clients. Figure 1 provides initial empirical support for the chosen interface, and Proposition 3.1 explains why improved aggregation fidelity alone cannot guarantee beneficial synchronization.

Critical weaknesses
  1. The evidence does not sufficiently connect aggregation fidelity to the method's performance gains or isolate its contribution. Equation (1) establishes a local linear property for fixed layer inputs, while Proposition 3.1 is a general risk comparison that does not specifically establish the superiority of value-bias adaptation or the proposed role assignment. Although Figure 1 reports lower normalized merge error, the diagnostic setup and controls for adaptation magnitude are insufficiently specified. Small updates can reduce this discrepancy even for nonlinear interfaces, so the comparison should account for the amount of useful adaptation achieved. Moreover, under the assumptions of Eq. (1), the injected value-bias correction can also be represented as an attention output-projection bias correction[1]. This makes output-projection bias tuning a particularly relevant missing baseline. The ablations also lack a role assignment control with matched state proportions and training/communication budgets. Consequently, the reported gains cannot yet be clearly attributed to the proposed fidelity argument and selection criterion rather than bias-only adaptation and partial synchronization more generally.

  2. The reliability of the one-time role assignment is not established for the subsequent training procedure. The sign of the probe loss difference determines a persistent role, yet the source, sampling procedure, sample count, and representativeness of the probe set are unclear; three probe batches may yield unstable decisions near the zero threshold. More fundamentally, the calibration score uses a global reference averaged over all clients, whereas later updates aggregate only the clients assigned to Shared for each group. Both the aggregation population and the surrounding model state therefore change. The initial Shared initialization does use the calibrated reference, but this does not establish the criterion's validity in later rounds. Appendix A analyzes joint replacements at the calibrated state; it does not establish transfer to the evolving selective aggregation procedure. Without evidence on probe stability and the persistence of sharing benefits, the central synchronization rule remains an insufficiently validated heuristic.

Questions for authors
  1. Aggregation mechanism and close baselines. What dataset, diagnostic samples, reference state, and training configuration were used for Figure 1? Can the authors provide a comparison with native attention output-projection bias tuning under comparable adaptation strength[1], reporting both functional change magnitude and merge error? Evidence relating these quantities to actual post-aggregation loss would materially strengthen the proposed explanation.

  2. Calibration versus actual synchronization. When the sharing score is recomputed using the actual selectively aggregated parameters in later rounds, how often do groups initially assigned to Shared still have non-positive replacement cost? Is there evidence that one-time calibration remains reliable compared with recalibration using the selected sharing clients or periodic role reassessment?

  3. Probe construction and selection-specific benefit. Is the probe set separate from the calibration training data, how is it sampled, and how many examples do three batches contain? How sensitive are role assignments and final accuracy to probe size and sampling seed? A random assignment control preserving each client's state proportions and resource budget would help determine whether the proposed criterion provides benefits beyond the allocation itself.

  4. Evaluation protocol and statistical reliability. In Table 4, do all domains participate in training, or is any domain held out? Please specify the client-level train/validation/test splits and whether evaluation uses personalized client models or a server model. How many independent runs were performed, and what is the variability of the reported results, particularly the 0.06-percentage-point average advantage on DomainNet-100?

Reference

[1] Baichuan Huang, Ananth Balashankar, and Amir Aminifar. BEFT: Bias-Efficient Fine-Tuning of Language Models in Low-Data Regimes. In Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 38833-38851, 2026.

2 comments
AuthorsOpenReview ↗

Clarifications on Aggregation Fidelity, Selective Synchronization, and Statistical Reliability

Thank you for the comments.

W1/Q1: Aggregation fidelity and output-bias comparison

Eq. (1) characterizes the additive correction of value-bias updates under fixed layer inputs, while Proposition 3.1 distinguishes aggregation fidelity from synchronization utility without guaranteeing performance superiority.

Figure 1 uses Stanford Cars with 10 clients and 8,041 fixed test images under the 16-shot protocol. We use SGD (LR 0.004, momentum 0.9, weight decay 0.0005) with five local epochs. Each method's global model before local training serves as the reference state. We evaluate the last five rounds (46–50) and report the median of 50 client-round merge errors.

We further compare value-bias and output-bias under matched functional displacement and local CE gain using 1,635 separate validation images:

MatchingInterfaceMatched ValueMerge ErrorΔL_merge
Functional displacementValue-bias0.3850140.0035800.000789
Output-bias0.3842520.0077330.001347
Local CE gainValue-bias0.0479000.0035800.000789
Output-bias0.0478460.0054120.001260

Here, ΔL_merge denotes the CE difference between parameter-averaged and averaged-logit predictions. Value-bias achieves lower merge error and post-aggregation CE deviation under both matching conditions. Under matched local CE gain, it also achieves lower actual post-aggregation CE (0.90775 vs. 0.91504 for output-bias).

W1/Q3: Matched-budget random assignment

We added a matched random-assignment baseline that preserves the Frozen/Shared/Private proportions for each client, encoder, and layer, with identical training and communication budgets.

MethodLast-10 Acc. (%)Trainable Params.Comm. (KB/client/round)
Matched Random78.94902411.264
TriVA (Ours)79.34902411.264

These results support the effectiveness of counterfactual role assignment beyond random allocation.

W2/Q2: One-shot calibration versus recalibration

We reevaluated initially Shared groups using actual selectively aggregated parameters in subsequent rounds. At rounds 10, 25, and 50, 62.05%–68.41% of these groups retain non-positive replacement costs.

We further compared one-shot calibration with selected-client recalibration and periodic role reassessment.

MethodLast-10 Acc. (%)Additional Calibration
Original TriVA79.34None
Selected-client79.386.32 min
Periodic Reassessment79.5797.52 min

Recalibration provides limited accuracy gains while introducing additional computation. One-shot calibration achieves comparable performance without repeated role reassessment.

W2/Q3: Probe construction and stability

Probe samples are drawn without replacement from each client's calibration training set, rather than a separate held-out set. Each batch contains 32 images, with 96 samples per client under the default three-batch setting. We evaluated sensitivity to probe size and sampling seeds on Stanford Cars.

SettingLast-10 Acc. (%)
1 batch79.25
3 batches (default)79.34
5 batches79.31
Probe seed 10179.30
Probe seed 10279.39
Probe seed 10379.32

Role agreement ranges from 50.33% to 55.21% across probe sizes and from 50.6% to 56.3% across sampling seeds.

Although individual role assignments vary, final accuracy remains relatively stable across the tested probe settings.

Q4: Evaluation protocol and statistical reliability

In Table 4, all four OfficeHome domains (8 clients) and six DomainNet-100 domains (12 clients) participate in training, with no held-out domains. Each domain is divided into two clients with disjoint training and test sets; no separate validation partition is used. Evaluation uses personalized client models, with accuracy averaged across clients within each domain and over the last 10 rounds.

We additionally report mean accuracy and standard deviation over three seeds on five single-domain benchmarks:

MethodCaltech101Food101UCF101CarsSUN397
CLIP93.16±0.2585.24±0.3165.08±0.4265.53±0.3962.58±0.33
PromptFL95.15±0.4483.76±0.7280.83±0.9177.54±0.7673.54±0.64
FedTPG95.06±0.2984.97±0.4979.95±0.5372.49±0.4771.02±0.51
FedOTP93.03±0.2493.87±0.3671.03±0.6262.51±0.5869.98±0.35
FedPHA93.38±0.2096.06±0.2678.01±0.4077.10±0.3472.46±0.31
pFedMMA94.71±0.2895.67±0.3379.33±0.4572.94±0.4171.15±0.38
pFedMoAP95.68±0.2695.93±0.3081.70±0.4775.17±0.4272.04±0.39
FedMGP96.08±0.3095.40±0.3883.67±0.5176.58±0.4474.16±0.45
TriVA96.26±0.3796.44±0.6882.76±0.3579.34±0.2974.75±0.56

Due to computational constraints, DomainNet-100 currently uses one training seed, and its multi-seed results will be included in the final version.

Reviewer_axJKOpenReview ↗

Reply to the authors

Thank you for the reply. The response only resolve part of my concerns. Especially, the reliability of the one-time role assignment is still questionable. Along with the prersentation issues. I will keep my original score.

Reviewer_k5PuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes TriVA, a personalized federated VLM adaptation method that tunes only value biases in the image and text encoders. A one-shot calibration assigns bias groups to Frozen, Shared, or Private states to balance collaboration and personalization. Experiments show competitive accuracy with low communication costs and no additional inference-time modules.

Critical strengths
  1. The paper provides a clear motivation for why value-bias updates are suitable for averaging.
  2. It combines shared and private adaptation with low communication costs.
  3. It tunes existing value biases without adding inference-time modules.
Critical weaknesses
  1. Since bias-based fine-tuning is well studied in the literature, I think the novelty comes from the calibration stage where the biased terms deiveied to three different groups. However, I have notable concerns regarding this stage. First, the empirical study for the calibration section is restricted to Stanford Cars, where removing Private improves accuracy and removing Shared causes a small decrease. So from the empirical point of view, the benefit of the calibration is not justified. Second, the role assignment during the calibration phase is inferred from three batches using zero thresholds, without evidence of their stability and validity during training. As a result, it is important for the paper to clearly broaden the ablations and comparison against the matched-budget random assignments and fully local adaptation.

  2. The evaluation in the paper focuses on personalization accuracy but does not assess base-to-new generalization, an important aspect of VLM adaptation evaluated by relevant baselines. Reporting base-class accuracy (accuracy over other clients), unseen-class accuracy, and their harmonic mean under a consistent protocol would clarify whether TriVA’s gains preserve pretrained generalization or primarily improve performance on locally observed classes.

  3. Some recent and strong baselines on personalization, such as pFedMoAP and FedMGP, are absent from the comparisons. Including these baselines provides a more comprehensive comparison.

  4. Table 5 covers only five datasets, leaving the gains on the remaining important benchmarks such as DTD, Oxford Pets, and Oxford Flowers evaluated by relevant baselines unclear. The paper also reports no seed counts or standard deviations, making it hard to tell whether the small improvements are meaningful.

  5. The paper lacks a sensitivity analysis of client count and participation rate. Varying these settings on the same dataset would help show whether TriVA and its fixed calibration assignments remain effective as the federation scales or fewer clients participate.

Questions for authors
  1. How stable are the calibration assignments across different probe batches and random seeds?
  2. Do the tri-state ablation findings in Table 7 hold beyond Stanford Cars?
  3. What is the total computation and communication overhead, including calibration?
  4. Does TriVA preserve base-to-new class generalization?
1 comment
AuthorsOpenReview ↗

Clarifications on Tri-State Calibration, Base-to-New Generalization, and Experimental Robustness

Thank you for the comments.

W1/Q2: Novelty and tri-state ablation comparisons

We agree that bias-based fine-tuning is well established. TriVA's novelty lies in its counterfactual role assignment within the intrinsic value-bias space, distinguishing adaptation utility from synchronization suitability to assign parameter groups to Frozen, Shared, or Private states.

We extended the tri-state ablation to SUN397 and included random assignment and local-only adaptation comparisons.

MethodLast-10 Acc. (%)
TriVA (Ours)74.75
Matched Random74.32
Fully Local74.14
w/o Frozen74.09
w/o Private74.39
All-Shared73.43

These results support the effectiveness of counterfactual role assignment and selective synchronization.

W1/Q1: Calibration assignment stability

We evaluated the stability of calibration assignments across different probe sizes and sampling seeds on Stanford Cars.

SettingLast-10 Acc. (%)
1 batch79.25
3 batches (default)79.34
5 batches79.31
Probe seed 10179.30
Probe seed 10279.39
Probe seed 10379.32

Role agreement ranges from 50.33% to 55.21% across probe sizes and from 50.6% to 56.3% across sampling seeds. During training, 62.05%–68.41% of initially Shared groups retain non-positive replacement costs at rounds 10, 25, and 50.

Although individual assignments vary, TriVA maintains consistent accuracy across the tested probe settings, indicating that its performance is relatively insensitive to these variations.

W2/Q4: Base-to-new generalization

We added base-to-new evaluations on five benchmarks under a consistent protocol, reporting base-class accuracy, novel-class accuracy, and their harmonic mean (HM).

DatasetBase (%)Novel (%)HM (%)
Caltech10197.4893.8295.61
Food10197.7585.6691.31
UCF10183.3072.1277.31
Stanford Cars70.7264.3667.39
SUN39775.0069.0871.92

Complete comparisons with all baselines are provided in Appendix Table 10.

W3: Additional personalized federated baselines

We added two recent personalized federated baselines, FedMGP and pFedMoAP, under the same federated evaluation protocol.

MethodCaltech101Food101UCF101Stanford CarsSUN397
FedMGP96.0895.4083.6776.5874.16
pFedMoAP95.6895.9381.7075.1772.04
TriVA (Ours)96.2696.4482.7679.3474.75

TriVA achieves the highest accuracy on four of the five benchmarks, demonstrating competitive performance against these additional baselines.

W4: Additional benchmarks and statistical reliability

We extended the add-on evaluation in Table 5 to DTD, Oxford Pets, and Oxford Flowers.

DatasetpFedMMA (%)pFedMMA + TriVA (%)
DTD68.24 ± 0.2268.29 ± 0.16
Oxford Pets93.50 ± 0.1693.58 ± 0.19
Oxford Flowers90.81 ± 0.2090.88 ± 0.27

TriVA provides consistent improvements across these additional benchmarks. The complete eight-benchmark add-on results are provided in Appendix Table 9.

We also report mean accuracy and standard deviation over three seeds for the single-domain benchmarks in Appendix Table 8. Due to computational constraints, multi-seed evaluations for the remaining settings will be included in the final version.

W5: Sensitivity to client count and participation rate

We evaluated the sensitivity of TriVA to client count and participation rate on Stanford Cars.

Client CountLast-10 Acc. (%)Participation RateLast-10 Acc. (%)
581.4120%78.91
10 (default)79.3450%79.70
2075.64100% (default)79.34

TriVA remains effective under different federation settings, with accuracy varying across client counts and remaining relatively stable under partial participation.

Q3: Total computation and communication overhead

We evaluated the computation and communication overhead of TriVA on Stanford Cars, including calibration. In runtime profiling, calibration takes 18.57 min (15.45 min for probing), accounting for 10.30% of the total 180.27 min runtime.

The cumulative communication over 50 rounds across 10 clients, including FP16 parameter payloads and metadata, is summarized below.

StageUplink (MB)Downlink (MB)Total (MB)
Calibration1.020.511.53
Rounds 2–504.654.659.30
Total5.665.1610.83

The total includes 6.13 MB of parameter payloads and 4.69 MB of metadata. The steady-state bidirectional payload is 11.26 KB/client/round.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.