acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Cross-Task Knowledge Sharing: Medical Federated Multi-Task Learning with the Input Space Knowledge Decoupling

Abstract

Federated multi-task learning (FMTL) seeks to share information across institutions and heterogeneous tasks while preserving client-specific personalization. Existing methods use local task models, their parameters, representations, or outputs, as carriers for federated knowledge sharing. Because potentially transferable and task-specific information are entangled within these carriers, aggregation cannot determine what should be shared and what should remain local: insufficient sharing may miss beneficial knowledge, whereas excessive sharing can cause negative transfer. To address this conflict, we propose IPL-FL, an FMTL framework that keeps personalization-specific knowledge local while separating cross-client knowledge sharing from local personalization, thereby avoiding the knowledge-selection conflict caused by their coupling. IPL-FL keeps task-specific knowledge in frozen local task models and instead communicates input-adaptation generators, preventing local predictors from being directly modified by aggregation. IPL-FL comprises client-side task-guided perturbation learning (TGPL) and server-side relation-aware personalized aggregation (RAPA). TGPL uses local task supervision to optimize task-beneficial input perturbations and transfers the resulting adaptations to shareable generators. RAPA decomposes generator updates into candidate shared information and client-specific residuals, and selectively aggregates compatible and complementary information using update-direction consistency and conditional information gain. Across eight clients spanning three medical tasks with heterogeneous architectures, IPL-FL improves every client over local training, achieves a 5.65% gain, exceeds the strongest applicable FMTL baseline by 5.27%, and offers a 13% efficiency advantage.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Multi-Task Learning, Medical Imaging, Cross-Task Knowledge Sharing

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_dT1TOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the knowledge-sharing problem in medical federated multi-task learning (FMTL), where transferable and task-specific knowledge are coupled within local task models, potentially causing negative transfer during aggregation.

The authors propose IPL-FL, a framework that decouples cross-client knowledge sharing from local task models through input-space perturbation generators, incorporating task-guided perturbation learning (TGPL) and relation-aware personalized aggregation (RAPA).

They evaluate the approach across eight clients spanning medical segmentation, classification, and detection tasks with heterogeneous model architectures, reporting a 5.65% average relative improvement over local training and outperforming existing FMTL baselines.

The study contributes an input-space knowledge-sharing framework for heterogeneous FMTL and an aggregation mechanism based on update-direction consistency and conditional information gain.

Critical strengths
  1. The paper identifies the knowledge-selection conflict in heterogeneous FMTL and provides empirical evidence of negative transfer through cross-task encoder aggregation experiments.
  2. IPL-FL separates task-specific modeling from federated knowledge sharing through TGPL and RAPA, enabling collaboration across heterogeneous tasks and model architectures without directly aggregating local predictors.
  3. The experiments cover eight clients across three medical task categories, with comparisons against heterogeneous FMTL methods and aggregation baselines demonstrating consistent improvements over local training and additional gains from RAPA.
Critical weaknesses
  1. The low-rank decomposition in RAPA assumes that shared subspaces capture transferable knowledge while residuals preserve client-specific information, but the experiments do not directly validate this distinction.
  2. Although the paper evaluates different perturbation configurations, it does not isolate the contributions of RAPA's low-rank decomposition, directional consistency, conditional information gain, and residual preservation
Questions for authors

see weaknesses

2 comments
AuthorsOpenReview ↗

Our Response to Reviewer dT1T

We thank the reviewer for your constructive comment.

A1: To empirically validate our knowledge decoupling assumption, we conducted an ablation study evaluating what happens when different decomposed components are aggregated across clients—specifically comparing aggregating only the residual component (EE), only the low-rank component (UpubrU_{\mathrm{pub}}^r), and both components (Upubr+EU_{\mathrm{pub}}^r + E):

a. Ablation of the low-rank component UpubrU_{\mathrm{pub}}^r and the residual component EE (Table 1)

ClientLocalLocal+TGPLEEUpubrU_{\mathrm{pub}}^rUpubr+EU_{\mathrm{pub}}^r+E
C10.75770.76840.75310.76950.7590
C20.35360.37080.36500.37860.3724
C30.82430.86490.85970.88490.8635
C40.93990.95320.95060.97090.9629
C50.80410.83620.81010.84580.8255
C60.80680.80940.79940.83610.8116
C70.27970.28330.29790.30940.2974
C80.49480.50450.50100.52670.5170
Δm\Delta m vs. Local (%)0.00+2.52+1.95+5.65+3.35

As shown in Table 1, aggregating only the residual component (EE) yields a Δm\Delta m of +1.95%, which is 0.57% lower than the non-aggregated Local+TGPL baseline (+2.52%). This performance degradation confirms that EE primarily captures task-specific information whose cross-client sharing causes negative transfer. In contrast, aggregating only the low-rank component (UpubrU_{\mathrm{pub}}^r) achieves the best performance (+5.65%), demonstrating that UpubrU_{\mathrm{pub}}^r isolates generalizable, transferable knowledge. Furthermore, although aggregating both components (Upubr+EU_{\mathrm{pub}}^r + E) improves Δm\Delta m by 0.83% over Local+TGPL (+3.35% vs. +2.52%), it still underperforms aggregating UpubrU_{\mathrm{pub}}^r alone by 2.30% due to interference from the task-specific residual. These findings directly validate our decoupling design.

Manuscript Revision: We have incorporated this component ablation study and analysis into Page 9 of the Main Paper.

A2: To isolate the individual contributions of all four components in RAPA, we conducted a comprehensive ablation study.

a. Ablation of the directional consistency (aija_{ij}) and conditional information gain (gj∣Sig_{j\mid S_i}) (Table 2)

ClientLocalLocal+TGPLaij=1, gj∣Si=1a_{ij}=1,\ g_{j\mid S_i}=1aij≠1, gj∣Si=1a_{ij}\ne1,\ g_{j\mid S_i}=1aij=1, gj∣Si≠1a_{ij}=1,\ g_{j\mid S_i}\ne1aij≠1, gj∣Si≠1a_{ij}\ne1,\ g_{j\mid S_i}\ne1
C10.75770.76840.76720.76900.76850.7695
C20.35360.37080.36900.37480.37270.3786
C30.82430.86490.85540.87210.87240.8849
C40.93990.95320.96590.96810.96360.9709
C50.80410.83620.82530.84470.83700.8458
C60.80680.80940.81140.82920.81960.8361
C70.27970.28330.29620.29630.29360.3094
C80.49480.50450.49890.51450.51610.5267
Δm\Delta m vs. Local (%)0.00+2.52+2.76+4.25+3.77+5.65

The results in Table 2 show that when both components are removed (aij=1,gj∣Si=1a_{ij}=1, g_{j\mid S_i}=1), performance drops to +2.76%, which is nearly identical to non-aggregated Local+TGPL (+2.52%). Between the two mechanisms, directional consistency exhibits a stronger individual impact: retaining only directional consistency (aij≠1,gj∣Si=1a_{ij}\ne1, g_{j\mid S_i}=1) achieves a Δm\Delta m of +4.25%, compared to +3.77% when retaining only conditional information gain (aij=1,gj∣Si≠1a_{ij}=1, g_{j\mid S_i}\ne1). This aligns with our intuition that filtering out conflicting knowledge is more critical than pruning redundant information in heterogeneous multi-task settings, while combining both mechanisms yields the full +5.65% improvement.

Manuscript Revision: We have incorporated directional consistency (aija_{ij}) and conditional information gain (gj∣Sig_{j\mid S_i}) ablation results into Page 9 of the Main Paper.

Thank you for the detailed rebuttal and additional experiments. I think the new ablation studies have addressed my concerns about the knowledge decoupling assumption and the contributions of different components in RAPA. Therefore, I will keep my original score unchanged.

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

This paper proposes IPL-FL, a federated multi-task learning framework that addresses knowledge-selection conflicts arising from the entanglement of transferable and task-specific knowledge in local models. Instead of directly sharing task-model parameters or representations, IPL-FL freezes local models and transfers knowledge through input-adaptation generators. The framework consists of TGPL, which learns task-beneficial input transformations, and RAPA, which selectively aggregates compatible and complementary knowledge across clients. Experiments across eight clients with heterogeneous medical tasks and model architectures demonstrate consistent improvements over local training and existing FMTL baselines.

Critical strengths
  • Clear presentation and well-organized methodology. The paper is well-written and easy to follow. The problem motivation is clearly presented, and the role of each proposed component in addressing specific challenges is well explained.

  • Practical and challenging problem setting. The paper considers a realistic federated learning scenario involving heterogeneous tasks, including segmentation, detection, and classification, as well as diverse model architectures such as YOLO, ResNet, and U-Net.

  • Detailed ablation studies. The paper provides several ablation studies examining perturbation optimization, generator architectures, and SVD configurations, helping to understand the contributions of different components.

Critical weaknesses
  • Unclear effectiveness of adaptive rank selection. The proposed method adaptively selects the minimum rank that preserves 90% of the singular-value energy, subject to a maximum rank of (qmax⁡=4q_{\max}=4). However, it is unclear how often the selected rank actually reaches this upper bound. If the rank is capped at 4 in most communication rounds, the method may effectively behave like fixed-rank truncation rather than adaptive energy-based selection. Moreover, the hyperparameter analysis shows considerable sensitivity to (qmax⁡q_{\max}), raising the question of whether the performance gains primarily stem from the 90% energy criterion or from choosing a favorable fixed rank. It would be helpful to report the distribution of selected ranks and the actual retained energy across communication rounds, as well as a comparison against fixed-rank SVD (e.g., always choosing 4 ranks).
  • Limited computational analysis. While the paper reports a 13% efficiency advantage over Ditto, it is unclear whether the reported training time includes the full computational cost. TGPL first optimizes perturbations for each training sample using 100 gradient steps and then trains a separate generator, potentially introducing substantial overhead. The authors should report end-to-end training time, including both stages, along with a computational cost breakdown to support their efficiency claims.
  • Insufficient comparisons with relevant baselines. Although the paper considers collaborative learning across heterogeneous tasks and model architectures, FedMosaic [1] also addresses knowledge sharing across different tasks, model sizes, and cross-family architectures (e.g., Llama vs. Qwen). Moreover, recent federated knowledge distillation methods, such as FedMKT [2], are missing from the baselines. - Comparisons with methods that selectively aggregate knowledge based on inter-client task similarity, such as NBS [3] and RELA [1], are also lacking. Given that the proposed framework consists of multiple components, more detailed comparisons with relevant baselines for each component would strengthen the empirical evaluation.
  • Insufficient validation of knowledge decoupling. The paper assumes that the low-rank component captures transferable knowledge while the residual preserves task-specific information. However, this assumption is not sufficiently validated empirically. A more detailed analysis of the information captured by each component is required
  • Limited scalability and generalizability evaluation. Both the main paper and appendix evaluate the proposed method using the same eight clients or subsets thereof. However, real-world federated learning systems may involve hundreds of clients spanning a broader range of tasks and data distributions. More extensive experiments with larger numbers of clients and diverse task combinations, potentially beyond the medical domain, are needed to demonstrate the scalability and generalizability of the proposed framework.

[1] Seo et al., Co-LoRA: Collaborative Model Personalization on Heterogeneous Multi-Modal Clients, ICLR 2026
[2] Fan et al., Fedmkt: Federated mutual knowledge transfer for large and small language models, COLING 2025
[3] Navon et al., Multi-task learning as a bargaining game, ICML 2022

Questions for authors

Please refer to the weakness section above.

4 comments
AuthorsOpenReview ↗

Our Response to Reviewer KnuY

We thank the reviewer for your constructive feedback.

A1: To thoroughly examine whether our method degenerates into fixed-rank truncation and to validate the 90% energy criterion, we tracked the selected rank distribution and retained singular-value energy across 100 federated communication rounds under the IPL-FL setting, and directly compared our adaptive strategy against fixed-rank truncation (r=4r = 4).

a. Rank Distribution and Retained Energy Statistics (Table 1):

Average rankProportion with rank 4Average energy
1.62.43%96.37%

As shown in Table 1, the average selected rank across 100 rounds is only 1.60, and the selection hits the upper bound (qmax⁡=4q_{\max}=4) in merely 2.43% of cases. This confirms that the rank selection operates genuinely in an adaptive regime rather than being capped at the upper bound. Moreover, the actual retained energy averages 96.37%, demonstrating that the 90% energy threshold effectively captures dominant principal components while allowing flexible rank adaptation across rounds.

b. Comparison Against Fixed-Rank Truncation (r=4r = 4) (Table 2):

ClientLocal$r=$4Adaptive rr
C10.75770.76740.7695
C20.35360.36190.3786
C30.82430.86550.8849
C40.93990.95390.9709
C50.80410.82600.8458
C60.80680.83160.8361
C70.27970.28840.3094
C80.49480.52750.5267
Δm\Delta m vs. Local (%)0.00+3.20+5.65

Table 2 demonstrates that our adaptive rank selection outperforms fixed-rank truncation (r=4r=4) on 7 out of 8 clients, improving the overall Δm\Delta m from +3.20% to +5.65%. While always fixing r=4r=4 retains more components, it inevitably includes minor singular components that carry client-specific noise, thereby degrading aggregation quality. In contrast, our 90% energy criterion adaptively filters out noisy tail components while preserving essential shared structure.

Manuscript Revision: We have incorporated these rank/energy statistics and the fixed-rank ablation study into Appendix A.7 and Pages 9–10 of the Main Paper.

A2: To clarify, Figure 4 of main paper evaluates the efficiency and performance during the federated aggregation phase. To fully address your concern, we provide a complete computational breakdown across both stages of IPL-FL below:

a. One-Time Sample Perturbation Optimization Time (Round 1 Only) (Table 3)

Optimize steps50100
Time / one image (ss)0.180.35
Δm\Delta m vs. Local (%)4.565.65

Perturbation optimization is performed only once during the initial round and is not repeated in subsequent communication rounds. As shown in Table 3, 100 optimization steps incur a one-time cost of 0.35s per image while boosting Δm\Delta m to +5.65% (or 0.18s per image with 50 steps for a +4.56% gain).

b. Generator Training vs. Local Backbone Training (Subsequent Rounds) (Table 4)

Training targetGeneratorLocal model (with U-Net)Local model (with TransUNet)
Time / one client (min)4.674.356.19
Parameters (M)9.326.5010.69

In subsequent rounds, only the lightweight prompt generator is updated while the local backbone remains frozen. Table 4 compares the per-round training time per client against standard FL methods based on training local model. First, with a compact U-Net backbone, generator training takes 4.67 min/client per round, closely matching standard local training (4.35min), while delivering substantial performance gains (+5.65%). Second, the cost of local-training FL scales directly with the backbone size, whereas IPL-FL's generator cost remains invariant to the local model architecture. Due to rebuttal time constraints, we verified this by replacing U-Net with TransUNet on clients C1 and C3: local backbone training time increased to 6.19 min/client per round, whereas IPL-FL remained unchanged at 4.67min. Thus, IPL-FL becomes even more computationally advantageous with larger, realistic backbones.

Manuscript Revision: We added this complete computational cost breakdown to Appendix A.6.

AuthorsOpenReview ↗

Our Response to Reviewer KnuY

A3:In response, we have added empirical comparisons against NBS and RELA in Table 2 of the revised manuscript, and clarified the fundamental paradigm differences regarding FedMKT and FedMosaic.

a. Quantitative Comparison with Similarity-Based Aggregation Baselines (NBS & RELA) (Table 5)

ClientLocalLocal+TGPLIPL-FL(Ours)NBSRELA
C10.75770.76840.76950.76370.7579
C20.35360.37080.37860.36800.3764
C30.82430.86490.88490.86820.8691
C40.93990.95320.97090.95790.9668
C50.80410.83620.84580.83940.8324
C60.80680.80940.83610.81160.8082
C70.27970.28330.30940.25040.2933
C80.49480.50450.52670.50770.5164
Δm\Delta m vs. Local (%)0.00+2.52+5.65+1.15+3.46

As shown in Table 4, IPL-FL outperforms both NBS and RELA across all 8 clients, achieving an overall Δm\Delta m of +5.65%, compared to +1.15% for NBS and +3.46% for RELA. Notably, on challenging clients such as C7, NBS suffers from negative transfer (0.2504 vs. 0.2797 for Local), whereas our RAPA mechanism effectively prevents interference and delivers consistent gains.

b. Fundamental Incompatibility with FedMKT and FedMosaic

While FedMKT and FedMosaic are notable cross-task/cross-architecture methods, direct experimental comparisons are infeasible due to fundamental differences in problem formulation and architectural assumptions.

For FedMKT: FedMKT relies on mutual knowledge distillation over a shared public dataset. In our heterogeneous multi-task setting, clients have vastly different output structures, task constraints, and semantic spaces, making it infeasible to construct a single public proxy dataset that satisfies all clients' task requirements. Adapting FedMKT to our setting would require redesigning its core distillation objective and cross-task output alignment interface. Therefore, forcing comparisons with FedMKT would depart from the core of the original method and compromise the fairness of the comparisons.

For FedMosaic: FedMosaic is designed for parameter alignment in Co-LoRA. Forcing modifications to its approach to accommodate the CNN or Transformer architectures used by IPL-FL would entail redesigning the method. Therefore, forcing comparisons with FedMosaic would depart from the core of the original method and compromise the fairness of the comparisons.

Manuscript Revision: We have incorporated the experimental results for NBS and RELA into Table 2 of the Main Paper, and cited FedMKT and FedMosaic in the Related Work.

A4: To empirically validate our knowledge decoupling assumption, we conducted an ablation study evaluating what happens when different decomposed components are aggregated across clients—specifically comparing aggregating only the residual component (EE), only the low-rank component (UpubrU_{\mathrm{pub}}^r), and both components (Upubr+EU_{\mathrm{pub}}^r + E):

a. Ablation of the low-rank component UpubrU_{\mathrm{pub}}^r and the residual component EE (Table 6)

ClientLocalLocal+TGPLEEUpubrU_{\mathrm{pub}}^rUpubr+EU_{\mathrm{pub}}^r+E
C10.75770.76840.75310.76950.7590
C20.35360.37080.36500.37860.3724
C30.82430.86490.85970.88490.8635
C40.93990.95320.95060.97090.9629
C50.80410.83620.81010.84580.8255
C60.80680.80940.79940.83610.8116
C70.27970.28330.29790.30940.2974
C80.49480.50450.50100.52670.5170
Δm\Delta m vs. Local (%)0.00+2.52+1.95+5.65+3.35

As shown in Table 6, aggregating only the residual component (EE) yields a Δm\Delta m of +1.95%, which is 0.57% lower than the non-aggregated Local+TGPL baseline (+2.52%). This performance degradation confirms that EE primarily captures task-specific information whose cross-client sharing causes negative transfer. In contrast, aggregating only the low-rank component (UpubrU_{\mathrm{pub}}^r) achieves the best performance (+5.65%), demonstrating that UpubrU_{\mathrm{pub}}^r isolates generalizable, transferable knowledge. Furthermore, although aggregating both components (Upubr+EU_{\mathrm{pub}}^r + E) improves Δm\Delta m by 0.83% over Local+TGPL (+3.35% vs. +2.52%), it still underperforms aggregating UpubrU_{\mathrm{pub}}^r alone by 2.30% due to interference from the task-specific residual. These findings directly validate our decoupling design.

Manuscript Revision: We have incorporated this component ablation study and analysis into Page 9 of the Main Paper.

AuthorsOpenReview ↗

Our Response to Reviewer KnuY

A5: Following the natural-image multi-task experimental setup of FedHCA$^2$ [1] and further incorporating the Cityscapes dataset, we conducted large-scale experiments involving 108 clients and 6 distinct tasks across three datasets (PASCAL-Context, NYUD-v2, and Cityscapes).

a. Large-scale Experiments (Table 7)

DatasetTaskNumber of clientsMetricLocalFedHCA2IPL-FL (ours)
PASCAL-ContextSemSeg4mIoU31.30 ± 13.1533.74 ± 12.5735.65 ± 6.48
PASCAL-ContextParts4mIoU34.63 ± 3.3733.64 ± 1.4234.79 ± 2.64
PASCAL-ContextNormals4mErr13.39 ± 1.0013.64 ± 0.8812.95 ± 1.32
PASCAL-ContextEdge4ODS65.72 ± 2.1267.47 ± 3.4467.96 ± 2.05
PASCAL-ContextSal4maxF74.92 ± 4.2372.87 ± 6.6175.84 ± 5.57
NYUD-v2SemSeg4mIoU25.73 ± 7.4525.84 ± 4.9027.77 ± 4.21
NYUD-v2Normals4mErr18.78 ± 1.3217.42 ± 2.6217.89 ± 3.81
NYUD-v2Depth4RMSE0.5159 ± 0.07370.5085 ± 0.08160.4933 ± 0.0362
NYUD-v2Edge4ODS73.60 ± 0.7275.26 ± 1.8775.92 ± 0.54
CityscapesSemSeg72mIoU47.83 ± 12.2445.71 ± 10.7749.79 ± 9.11
Δm\Delta m vs. Local (%)---0.00+0.99+4.66

Table 7 summarizes the client distribution per task and compares Local, FedHCA$^2$, and IPL-FL. Even at a scale of 108 clients across diverse natural-image tasks, IPL-FL maintains strong performance, achieving an overall Δm\Delta m of +4.66% compared to only +0.99% for FedHCA$^2$ (which suffers from negative transfer on the 72-client Cityscapes task). These results demonstrate the strong scalability and cross-domain generalizability of our framework.

[1] FedHCA$^2$: Towards hetero-client federated multi-task learning[C], CVPR, 2024.

Manuscript Revision: We have incorporated these large-scale natural-image experimental results into Page 10 of the Main Paper.

AuthorsOpenReview ↗

Our Response to Reviewer KnuY

Dear Reviewer KnuY

Thank you again for your review and insightful suggestions. We have uploaded a revised PDF (changes are marked in blue), which adds baselines NBS and RELA (Table 2), a discussion of FedMKT and FedMosaic (Related Work), component-wise ablations decoupling UpubrU_{\mathrm{pub}}^r and EE (Page 9), adaptive rank/energy statistics and fixed-rank (r=4r=4) comparisons (Pages 9–10, Appendix A.7), large-scale experiments across 108 clients on PASCAL-Context, NYUD-v2, and Cityscapes (Page 10), and a two-stage computational cost breakdown across U-Net and TransUNet (Appendix A.6). Could you let us know whether any of your concerns remain unresolved? We would be glad to address them before the discussion period ends.

Reviewer_mPhfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes IPL-FL for medical federated multi-task learning with heterogeneous predictors. Task-guided perturbation learning (TGPL) optimizes bounded, label-supervised input modifications and trains a generator to reproduce the optimized images. Relation-aware personalized aggregation (RAPA) decomposes generator updates into a low-rank component and a retained residual, constructs a graph using directional alignment and conditional information gain, and smooths only the low-rank coordinates. Experiments on eight clients spanning classification, segmentation, and detection report a 5.65% mean relative improvement over local training.

Critical strengths

The knowledge carrier offers a practical route around predictor heterogeneity. A common image-adaptation generator can be communicated even when the local classification, segmentation, and detection networks have incompatible parameter spaces. Keeping the predictors fixed makes the location of federated modification explicit.

Critical weaknesses

The headline comparison does not reflect the strongest control in the main results. The reported 5.27 figure is the difference from FedMuscle in Table 1. In Table 2, however, ConFREE reaches 5.25% mean relative improvement, compared with 5.65% for IPL-FL. This 0.40% margin is more relevant to assessing the proposed aggregation mechanism, and no repeated-run uncertainty is reported. The positive comparison against Local+TGPL is useful, but the paper should distinguish evidence that RAPA is superior to strong aggregation alternatives.

Questions for authors

Can the authors report repeated-run uncertainty for the comparisons in Table 2, particularly IPL-FL versus ConFREE?

4 comments
AuthorsOpenReview ↗

Our Response to Reviewer mPhf

We thank the reviewer for your constructive feedback.

A1: We have prioritized the comparison with ConFREE, clarified the specific contribution of our aggregation mechanism (RAPA), and quantified performance uncertainty across repeated runs. Specifically, we conducted 10 independent runs with different random seeds for each method and report the mean / sample standard deviation in the table below:

a. Uncertainty analysis of the method (Table 1)

ClientLocalLocal+TGPLIPL-FL(Ours)ConFREE
C10.7595/0.01210.7675/0.00750.7705/0.00890.7696/0.0114
C20.3521/0.00460.3741/0.01180.3800/0.00680.3793/0.0041
C30.8328/0.01650.8722/0.00660.8892/0.01130.8812/0.0016
C40.9374/0.01110.9539/0.00710.9699/0.00550.9636/0.0062
C50.8064/0.00680.8420/0.01520.8554/0.01130.8400/0.0071
C60.8098/0.01350.8115/0.00180.8386/0.00840.8260/0.0087
C70.2812/0.00710.2868/0.00370.3138/0.01120.3167/0.0187
C80.4953/0.00350.5053/0.01650.5321/0.00980.5145/0.0115
Δm\Delta m vs. Local (%)0.00+2.80+6.03+5.04
Δm\Delta m vs. Local+TGPL (%)-0.00+3.16+2.19

Across the 10 runs, IPL-FL achieves a Δm\Delta m of +6.03% relative to the Local baseline, outperforming ConFREE (+5.04%) by a clear margin of +0.99%. Furthermore, to isolate the effectiveness of federated aggregation from local prompt learning (Local+TGPL), we compute Δm\Delta m relative to Local+TGPL: our RAPA aggregation achieves +3.16%, compared to +2.19% for ConFREE. These statistically grounded results confirm that RAPA provides a consistent advantage over strong aggregation alternatives.

Manuscript Revision: Due to time constraints during the rebuttal period, 10-seed evaluations for the remaining baselines (FedAvg, FedAdam, NBS, FedAWA, FedMerge, and RELA) are currently in progress and will be thoroughly incorporated into Appendix A.5.

Thank the authors for the additional experiments.

I remain concerned about the performance comparison. While IPL-FL outperforms ConFREE by 0.99%, the difference in mean performance is smaller than standard deviation on many clients. So it remains unclear whether the average advantage is statistically significant.

AuthorsOpenReview ↗

Our Response to Reviewer mPhf

Thank you for this important clarification. We agree that higher mean performance and individual standard deviations alone do not establish statistical superiority, and our previous wording overstated the evidence.

In response, we have now evaluated the paired difference in overall mean relative improvement, treating each complete eight-client federated run as one independent experimental unit. Relative improvements are calculated separately for each run before averaging. The results are:

Reference for relative improvementMean difference: IPL-FL − ConFREE (%)95% confidence intervalTwo-sided paired t-test p-value
Local (primary comparison)+0.9888[0.3415, 1.6361]0.006730
Local+TGPL (secondary analysis)+0.9610[0.3315, 1.5905]0.006752

For the primary comparison, the 95% confidence interval excludes zero, with t(10)=3.4036t(10) = 3.4036. An additional exact sign-flip test yields p=0.0078125p = 0.0078125, supporting the same conclusion under its symmetry assumption. These results support a statistically significant aggregate advantage in the evaluated setting. They do not establish significant superiority on every client, and we have accordingly moderated our earlier claim of a consistent advantage.

Thank you for the response. I remain concerned about the effectiveness of the proposed method, as the improvements are marginal and exhibit considerable variance. These results also raise concerns about the method's generalizability to broader scenarios. Therefore, I will maintain my original score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Gratitude to the PC, SAC, AC and All Reviewers for Your Time and Efforts

We sincerely thank the PC, SAC, AC and all reviewers for your valuable time and insightful feedback. Guided by the reviewers' suggestions, we have substantially revised the manuscript and added targeted experiments to strengthen the empirical validation of IPL-FL and its design choices. We kindly invite the Area Chair to review our updated manuscript and individual responses. If these additions satisfactorily resolve the initial concerns, we respectfully request a positive reconsideration of our work's assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.