acceptodds
Under review as a conference paper at ICLR 2027

Dual-Partial Multimodal Federated Learning via Uncertainty-Gated Modality–Label Evidence Graphs

Abstract

Multimodal federated learning (MFL) becomes particularly difficult when modalities and labels are both partially observed: missing modalities make cross-modal transfer necessary, whereas missing labels remove the supervision needed to verify whether that transfer is correct. This evidence dilemma makes the most transfer-dependent targets the least directly verifiable. We formulate dual-partial MFL as progressively verified evidence transfer and propose Dual-Partial Evidence-Graph Federated Learning (DualP-EGFL). DualP-EGFL constructs signed modality–label evidence prototypes from verified observations, admits a cross-modal relation only after checking structural support and verified target-side utility, and applies accepted relations to an unknown target only when pair-specific out-of-distribution (OOD) gating and conservative Top- path aggregation provide sufficient support. Unsupported targets remain low-evidence rather than receiving fabricated pseudo-labels. Across benchmark settings, DualP-EGFL improves supported-hidden Macro-AUPRC by up to 7.65 percentage points over recent MFL baselines, with the largest gain in the 100-client MM-IMDb setting; in calibrated smaller-client settings, it also reduces false support relative to ungated graph transfer.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Federated Learning, Multimodal

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_kGY9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DualP-EGFL, a framework for multimodal federated learning under jointly missing modalities and labels. The method constructs modality–label evidence graphs from verified observations and introduces a two-level verification mechanism to improve the reliability of cross-modal knowledge transfer. Specifically, edge-level verification evaluates structural compatibility and predictive utility, while sample-level verification employs OOD gating and Top-K path aggregation to selectively provide supervision for missing labels. Unsupported labels are regularized to maintain low evidential confidence. The authors evaluate the proposed method on MM-IMDb, MELD, and synthetic datasets, including a 100-client federated setting. Experimental results demonstrate improvements in hidden-label prediction over existing baselines, supported by component ablations and transfer reliability analyses.

Critical strengths
  1. The paper addresses the challenging and practically relevant setting of multimodal federated learning with jointly missing modalities and labels. By highlighting the difficulty of verifying cross-modal knowledge transfer under incomplete supervision, the work identifies an important problem that is not sufficiently addressed by existing approaches.

  2. The proposed method achieves notable improvements over existing baselines, particularly in the 100-client MM-IMDb setting, where supported-hidden Macro-AUPRC improves by 7.65 percentage points over PePSY. The additional component ablations and transfer reliability analyses provide useful insights into the proposed framework, supporting its potential effectiveness in recovering missing supervision.

Critical weaknesses
  1. Although reliable cross-modal knowledge transfer is the central motivation of the paper, the experimental results raise serious concerns about the effectiveness of the proposed verification mechanism. In the 100-client MM-IMDb setting, the false-support rate reaches 100% with 99.99% transfer coverage, indicating that the method fails to reject unsupported transfers. Conversely, the low false-support rate of 0.68% on MELD is achieved with only 3.76% coverage. These results reveal a substantial trade-off between transfer reliability and coverage, undermining the paper's central claim of reliable knowledge transfer.

  2. The proposed framework primarily combines established techniques, including prototype learning, Procrustes alignment, uncertainty estimation, OOD detection, and Top-K aggregation. Although their integration is motivated by the dual-partial setting, the paper does not sufficiently establish what fundamentally new mechanism emerges from this combination. Moreover, the ablation studies do not adequately justify several central design choices, particularly target-side predictive utility verification and Top-K path aggregation. This makes it difficult to determine whether the complexity of the proposed framework is necessary or whether simpler alternatives could achieve comparable performance.

Questions for authors
  1. DualP-EGFL involves multiple interconnected components, including signed modality–label evidence graph construction, edge-level transfer verification, sample-level OOD gating, and evidential knowledge transfer. However, the paper lacks an overall architecture or workflow diagram illustrating how these components interact within the federated learning framework. As a result, readers may find it difficult to quickly understand the complete training procedure and the information exchanged between clients and the server.
  2. A central contribution of this paper is to prevent unreliable cross-modal knowledge transfer through edge-level and sample-level verification. However, Section 5.4 reports that the 100-client MM-IMDb experiment achieves 99.99% transfer coverage with a 100% false-support rate (FSR). This indicates that the proposed verification mechanism fails to reject unsupported transfers in the large-scale setting, despite the substantial improvements in prediction performance. Such a result directly challenges the paper's central claim of reliable knowledge transfer.
  3. On MELD, DualP-EGFL achieves a low FSR of 0.68%, but its transfer coverage is only 3.76%, meaning that almost all hidden entries are rejected from receiving transferred supervision. In contrast, the six-client MM-IMDb setting achieves 75.41% coverage but still suffers from a 55.01% FSR. These results raise concerns about whether the proposed mechanism provides a practically useful balance between rejecting unreliable transfers and recovering missing supervision. The authors should report coverage–reliability trade-off curves across datasets and demonstrate meaningful improvements at comparable coverage levels.
  4. Equation (13) defines path reliability as (g_j=r_{\mathrm{pair}}\exp[-\min(d_j^+,d_j^-)/\tau_j]), which depends on the distance to the nearest positive or negative prototype rather than the separation between them. Consequently, a sample with (d_j^+\approx d_j^-\approx0) may receive high reliability even though Equation (10) produces (p_j^{\mathrm{tr}}\approx0.5), indicating little discriminative evidence. This suggests that the gating mechanism may confuse proximity to the prototype distribution with confidence in the transferred prediction. The authors should justify this design and compare it with margin-aware or uncertainty-aware applicability scores.
  5. The paper uses false-support rate (FSR) as a key metric to demonstrate the reliability of its knowledge transfer mechanism. However, it is unclear how the ground-truth support status of hidden entries is independently established, particularly on real-world datasets such as MM-IMDb and MELD. Since a hidden entry may have an unknown label but still be supported by useful cross-modal evidence, support cannot be determined solely from label observability or the model's own gating decision. The authors should explicitly describe how supported and unsupported targets are identified for evaluation and demonstrate that this definition is independent of the proposed verification mechanism.
  6. Proposition 1 establishes the concentration bound (|\hat{u}-u^\star|\leq\frac{5}{2}\sqrt{\log(8/\delta)/(2n_{\min})}), assuming that the calibration samples are independent of the predictor being evaluated. However, the proposed framework constructs cross-modal prototypes using verified observations and subsequently evaluates transfer utility on verified target-modality samples. The paper does not clearly specify whether these procedures use disjoint data subsets. If the same observations are reused for prototype construction and utility verification, the independence assumption may be violated, and the concentration bound would not directly apply. The authors should clarify whether held-out calibration or cross-fitting is implemented.
  7. The proposed bound is (|\hat{u}-u^\star|\leq\frac{5}{2}\sqrt{\log(8/\delta)/(2n_{\min})}), where (n_{\min}) denotes the smaller number of verified positive and negative calibration samples. At (\delta=0.05), the bound exceeds 1 when (n_{\min}=10), making it vacuous since the transfer utility is bounded within ([0,1]). Achieving a bound below 0.1 would require approximately 1,586 verified samples from each class. This requirement may be difficult to satisfy in the very dual-partial setting considered by the paper. The authors should report the actual calibration sample counts and quantify how often their theoretical guarantee is non-vacuous.
  8. The theorem assumes that supported targets have at least (K) paths satisfying (r_j\geq r_{\mathrm{in}}), (d_j\leq d_{\mathrm{in}}), and (\tau_j\geq\underline{\tau}), while unsupported targets satisfy (r_j\leq r_{\mathrm{out}}), (d_j\geq d_{\mathrm{out}}), and (\tau_j\leq\overline{\tau}). Under these assumptions, a separating threshold exists when (Kr_{\mathrm{out}}e^{-d_{\mathrm{out}}/\overline{\tau}}<\tau_g\leq Kr_{\mathrm{in}}e^{-d_{\mathrm{in}}/\underline{\tau}}). However, the theorem does not explain why the learned evidence graph should satisfy these separation conditions, particularly under severe modality and label missingness. The authors should provide empirical evidence of such separation and discuss how often the conditions hold in practice.
1 comment
AuthorsOpenReview ↗

Response to Reviewer kGY9

W1: Reliability-Coverage Failure

Response: Our mechanism achieves its design goal—controllable rejection of unreliable transfers—which prior methods lack. Utility verification reduces MM-IMDb-100 coverage from 99.90% to 5.41%, cutting FSR from 99.83% to 13.15% and teacher error from 26.50% to 16.34%. At 5% coverage: FSR 0.71% (IMDb-6), 10.14% (IMDb-100), 5.04% (Synthetic). Despite conservative rejection, DualP-EGFL achieves +10.41 pp gain over FedAvg on MM-IMDb-100 [95% CI: +9.18, +11.64].

Extreme points clarified: MELD 0% coverage reflects insufficient cross-modal evidence, not failure; Synthetic low coverage reflects correct rejection under adversarial signals. Our contribution is threefold: (1) formulate when to transfer as an explicit decision, (2) provide dual-level verification for control, and (3) expose the coverage-reliability trade-off prior methods ignore.

W2: Established Components and Insufficient Ablations

Response: Revised ablations isolate utility checking, Top-K vs unrestricted sum, applicability type (distance/margin/entropy), OOD gating, and graph distillation. MM-IMDb-100 without utility: AUPRC 54.06% vs 53.73%; Synthetic unrestricted sum: 98.20% vs 98.07%. Revised claim: This is an integration study with explicit rejection trade-offs, not a proof each component improves prediction. Results in Table 2 and Appendix Table 8.

Q1: Overall Workflow

Response: Framework figure (revised manuscript Page 2) shows complete workflow and information exchanges. We clarify this remains centralized simulation, not secure distributed deployment.

Q2: MM-IMDb-100 Accepts Unsupported Transfers

Response: With utility verification: Coverage 5.41%, FSR 13.15%, Teacher Error 16.34%. Without verification: Coverage 99.90%, FSR 99.83%, Teacher Error 26.50%. Dual-level verification reduces false acceptance by 86.68 pp and teacher error by 10.16 pp.

Q3: Curves at Similar Coverage

Response: We provide curves at 6 target coverage levels with achieved coverage, FSR, teacher error, and Brier score. Complete per-seed curves in supplementary CSVs. Results in Appendix Table 7.

Q4: Prototype Proximity vs Discriminative Confidence

Response: Confidence-aware applicability (margin/entropy) improves teacher correctness by 9.79-13.42 pp and Brier score by 0.042-0.063 at similar coverage on MM-IMDb-6, validating this concern. However, both alternatives reject everything (0% coverage) on MM-IMDb-100, MELD, and Synthetic at default thresholds, requiring dataset-specific retuning. We chose distance-based applicability for stable generalization but acknowledge incorporating discriminative confidence as promising future work. Results in Appendix Table 11.

Q5: Independent Support Ground Truth

Response: Support truth is computed independently from available modalities, protocol allowed-transfer tensor, and OOD flags—not from predictions or gate outputs. Real-data flags do not physically perturb features; these are protocol support metrics, not natural OOD detection measures. MM-IMDb-6 default: protocol precision 96.83±2.12%, teacher error 29.84±6.78%, confirming these are distinct metrics.

Q6: Independence Assumption in Proposition 1

Response: Revised runner reserves disjoint optimization (70%), development (15%), audit (15%) splits. Normalization uses optimization rows only; final calibration performed once after fitting. Row counts: MM-IMDb-6 (1,008/216/216), MM-IMDb-100 (8,400/1,800/1,800), MELD-6 (~5,122/1,096/1,099), Synthetic (1,134/243/243). Addresses direct reuse but not correlated observations. Claim is pointwise for audit-independent frozen predictor, not simultaneous certification.

Q7: Vacuous Concentration Bound

Response: At δ=0.05, bound radius ~1.26 at n_min=10 (vacuous). Radius <0.1 requires ≥1,586 samples/class. Median n_min: 13.5 (Synthetic) to 111.5 (MELD-6). No pair achieves radius <0.1 across datasets. Results do not support tight guarantees in sparse regimes. Analysis in Appendix Table 14.

Q8: Empirical Support-Separation

Response: Training does not guarantee Theorem 1 separation. For K=2: supported samples have ≥2 paths in 11.78-73.86% of cases; unsupported samples have ≥2 paths in 1.01-74.14%. None of 19 runs achieve perfect separation. Theorem remains explanatory conditional result, not empirical gating justification. Results in Appendix Table 15.

Reviewer_wuKWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Dual-Partial Evidence-Graph Federated Learning (DualP-EGFL) by framing multimodal federated learning (MFL) as progressive verified evidence transfer to address two key challenges: reliable relation discovery under sparse verified overlap and selective supervision for unverifiable targets.

To solve the first challenge, DualP-EGFL constructs signed modality-label evidence prototypes from verified observations and accepts candidate relations only after confirming structural support and predictive utility in verified target modality data. To solve the second challenge regarding unknown targets, it calibrates pair-specific out-of-distribution (OOD) measures, scores sample-level path reliability, and aggregates only the most strongly supported paths. By formulating an evidence dilemma in MFL as reliable knowledge transfer and evaluating the framework across 100 clients under a joint modality and supervision incompleteness environment, the paper demonstrates superior performance compared to recent MFL baselines.

Critical strengths
  1. The study reframes MFL around the core question of "how to safely transfer reliable evidence/knowledge," effectively defining it as a knowledge transfer problem. Unlike prior MFL studies that address missing modalities using simple feature alignment, imputation, or contrastive learning to cluster text and image vectors closely together, this work offers a refreshing and distinct perspective.
  2. The evaluation under a challenging joint-modality and supervision-incompleteness environment reflects practical, real-world constraints, making the research highly relevant to realistic FL scenarios.
Critical weaknesses
  1. There is a notable absence of sensitivity analysis for hyperparameter settings. For better reproducibility, detailed explanations of hyperparameters and empirical validation of their variations are necessary. Inconsistent terminology and a lack of verification for key hyperparameters undermine the overall reliability.
  2. To enhance the readability and clarity of the proposed methodology, a framework figure is missing. While the mathematical details are thorough, adding a visual overview of the method would greatly improve comprehension.
  3. The performance margins over existing baselines do not always appear distinct, with several cases showing comparable results. Including standard deviations in the experimental results is strongly recommended to substantiate the statistical significance of the improvements.
Questions for authors
  1. The abstract mentions "Conservative Top-K," whereas the main text refers only to "Top-K." What is the exact value of K used in the experiments, and could you provide a performance graph showing the impact of varying K?
  2. The manuscript does not specify the non-IID data partitioning setting (e.g., the Dirichlet distribution hyperparameter). Was non-IID data distribution considered, and if so, how was it parameterized across clients?
  3. How was the threshold for Uncertainty-Gating determined?
  4. When clients exchange signed evidence, how do you address or mitigate potential privacy leakage risks?
1 comment
AuthorsOpenReview ↗

Response to Reviewer wuKW

W1: Missing Sensitivity Analysis

Response: We completed 28 one-factor sensitivity settings per dataset (5 seeds each). Key hyperparameters are K=2, τ_e=0.18, τ_g=0.10-0.65 (dataset-dependent), utility floor=0.45-0.50, λ_trans=0.035-0.12, and λ_unc=0.01-0.02. For τ_g ∈ {0.012, 0.024, 0.05, 0.10, 0.20}, AUPRC varies minimally (54.50-54.88% on MM-IMDb-6) while coverage varies greatly (1.38-39.04%), confirming threshold primarily controls rejection behavior rather than predictive performance. For K sensitivity, two-modality datasets show identical results across K=2/3/4/6, while three-modality Synthetic shows increasing coverage (0% to 22.31%) with minimal AUPRC change (98.07% to 98.21%) as K increases from 1 to 6. Complete sensitivity results are in Appendix Table 9.

W2: Missing Framework Figure

Response: We have created a comprehensive workflow figure showing the complete dual-level verification process, available at revised manuscript Page 2 (Figure 1).

W3: Small Margins and Missing Standard Deviations

Response: All 5-seed results now include mean ± SD. MELD gain over FedAvg is +0.66 pp with 95% CI [-0.32, +1.96] (statistically uncertain). MM-IMDb-6 shows CreamFL and PePSY outperform DualP-EGFL. We remove "uniformly best" language from the manuscript.

Q1: Exact K and Its Effect

Response: K=2 is used consistently in revised experiments. With two modalities, K=1 vs K=2 changes coverage (6.72→8.25% on MM-IMDb-6) but not AUPRC. We describe K=2 as an explicit conservatism control, not an empirically optimal setting.

Q2: Non-IID Partitioning

Response: Yes—all experiments use non-IID partitioning. MM-IMDb-6/100 use primary-genre-skewed sampling with strengths 0.70/0.75; MELD-6 uses realistic label/speaker/season skews 1.8/1.5/1.2. Non-IID sensitivity on MM-IMDb-6 shows AUPRC increasing from 42.22% (IID) to 61.92% (skew=0.9), with higher skew also changing task composition. Results are in Appendix Table 10.

Q3: Uncertainty-Gating Threshold

Response: There is no separate "Uncertainty-Gating" threshold. Edge uncertainty contributes to structural reliability (filtered by τ_e); τ_g controls sample support. Development threshold selection using prespecified diagnostic (max coverage with ≤10% development teacher error, ≥30 accepted entries) found no feasible threshold for MM-IMDb-6 and MELD-6 in any of 5 seeds. This diagnostic is adaptive and must not be presented as a pre-audit deployment procedure or certified error bound.

Q4: Privacy of Signed Evidence

Response: Addressed in R1-Q3. "Signed" means positive/negative label evidence, not cryptographic authentication. Membership AUROC remains ~0.69-0.76 even with count suppression.

Reviewer_Q6M7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This This paper focuses on multimodal federated learning in the presence of both missing modalities and missing labels. The authors point out that existing methods generally concentrate on enabling knowledge sharing among heterogeneous clients, yet fail to identify when cross-modal knowledge should be transferred. To address this issue, the authors propose DualP-EGFL, which constructs a modality-label evidence graph. Integrating multiple sources of information such as prototype similarity and sample uncertainty, this graph filters modality relationships worthy of transfer. A gating mechanism is further adopted to determine which samples receive knowledge transfer. The authors conduct experiments on multiple datasets and achieve performance superior to the baselines.

Critical strengths
  1. The paper presents a valuable problem motivation, and its method design is reasonably sound.
  2. The paper includes some theoretical analysis.
Critical weaknesses
  1. The methods adopted in this paper are mostly existing techniques. Although the authors combine them to tackle a new problem, the work does not introduce substantial novelty at the methodological level.

  2. All experiments in the paper are conducted under a single random seed. This may introduce substantial randomness into the results, making it difficult to verify the robustness of the proposed method. Federated learning is highly sensitive to data partitioning and algorithm hyperparameters. Therefore, we recommend running the main experiments with at least five random seeds and reporting the mean values.

Questions for authors
  1. According to the experimental results, does the proposed method not seem to achieve high transfer coverage and reliability simultaneously?
  2. Why is the FSR 100% under the 100-client setting of MM-IMDb? This seems to suggest that, in some cases, the gating mechanism proposed in this paper is completely ineffective.
  3. This paper does not appear to include systematic analysis or even simple experimental validation regarding the privacy security of the proposed method. Could the authors provide this part?
  4. Since the proposed method constructs a relatively complex graph, analysis of the computational complexity is particularly important.
  5. Could the authors construct a baseline method based on prototype transfer, vanilla gating and confidence filtering to justify the necessity of the evidence graph design?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Q6M7

W1: Limited Methodological Novelty

Response: No—while individual components exist, our contribution is not merely incremental combination.

Our Novelty: DualP-EGFL addresses a problem that existing techniques do not solve individually: reliable knowledge transfer when both modalities and labels are jointly incomplete. This creates the evidence dilemma—targets needing transfer most are least verifiable. Our contribution lies in formulating and solving this dilemma through dual-level verification:

  1. Edge-level verification determines whether a cross-modal relation preserves target decision structure using verified target-modality data (not just geometric compatibility)
  2. Sample-level verification determines which unknown targets should receive transferred supervision via pair-specific OOD calibration and conservative Top-K aggregation

What Makes This Unique:

  • Prior prototype methods (FedProto, FedTGP) transfer based on geometric similarity alone without target-side utility validation
  • Prior reliability methods (NucFL, FedMDMI) calibrate prediction confidence but not transfer applicability
  • DualP-EGFL uniquely separates structural plausibility from target-discriminative utility and applies verified relations only to in-distribution targets

Ablation Evidence for Dual-Level Design: We constructed the requested "Prototype + Confidence" baseline using signed prototypes, alignment, and 0.8 probability cutoff—bypassing both utility verification and sample-level gating. The simple baseline achieves comparable AUPRC (54.39% vs 54.57% on MM-IMDb-6) but dramatically different reliability: MM-IMDb-6 FSR increases from 1.16% to 55.20% (47× higher), and Synthetic coverage increases from 0.019% to 11.68% with FSR rising from 0.035% to 13.80%. The comparable AUPRC validates that prediction accuracy is orthogonal to reliability control—our dual-level verification controls when to transfer, not just how to predict.

W2: Single-Seed Evaluation

Response: We have completed 5-seed main comparisons with common partitioning and missingness protocols. DualP-EGFL achieves 54.57±4.01%, 53.73±1.29%, 30.79±3.16%, and 98.07±2.15% Supported-Hidden Macro-AUPRC on MM-IMDb-6, MM-IMDb-100, MELD-6, and Synthetic respectively, with largest gains in the 100-client setting (+10.41 pp over FedAvg). Full results are provided in Appendix Table 6 (Multi-seed Main Results).

Q1: Can Coverage and Reliability Be High Simultaneously?

Response: Yes—but the operating point depends on threshold choice, revealing a fundamental coverage-reliability trade-off that our method explicitly controls. At default thresholds, MM-IMDb-6 achieves 8.25% coverage with 1.16% FSR, MM-IMDb-100 achieves 5.41% coverage with 13.15% FSR, while MELD-6 shows 0% coverage (complete rejection due to insufficient verified evidence). At 5% target coverage, all datasets achieve single-digit FSR (0.71-10.14%), demonstrating tunable trade-offs. Complete curves at 6 coverage levels (1%, 5%, 10%, 25%, 50%, 75%) are provided in Appendix Table 7 (Coverage-Reliability Trade-off).

Q2: Why Is MM-IMDb-100 FSR 100%?

Response: The historical result had an incomplete configuration. Our revised protocol with full dual-level verification achieves FSR 13.15±6.21%—an 86.68 percentage point improvement. Without utility checking, coverage reaches 99.90% with FSR 99.83%, confirming that utility verification is essential and effective.

Q3: Privacy and Security Validation

Response: We clarify that keeping raw data local is not a formal privacy guarantee. We conducted honest-but-curious server probes using class-conditioned centroid-distance membership ranking. Membership inference AUROC ranges from 0.695±0.032 (Synthetic) to 0.767±0.044 (MM-IMDb-6). Increasing minimum count suppression from 1→10 modestly reduces leakage but does not eliminate it. We explicitly state that raw-input inversion, model-update leakage, secure aggregation, and differential privacy remain untested/unimplemented. Full results are in Appendix Table 12.

Q4: Computational Complexity

Response: Theoretical complexity includes structural scoring O(M²Cd), calibration O(P·N_cal·d), and Top-K selection O(B·C·P·log P). Empirical measurements show DualP-EGFL is approximately 2.5× slower than FedAvg in centralized simulation (scaling from 2.19× at 6 clients to 2.56× at 100 clients). Memory overhead includes 16.2±0.9 MiB for moments and 73.5 MiB for graph structures. Detailed scaling results are in Appendix Table 13 (Computational Complexity).

Q5: Simpler Prototype/Gating/Confidence Baseline

Response: Addressed in W1. The simple baseline validates our dual-level design principle by achieving comparable prediction accuracy but failing at reliability control, confirming our contribution is orthogonal to predictive performance.


Thank the authors for the detailed rebuttal and additional experimental results. I would maintain my original score. The authors have substantially improved the reliability analysis, but the rebuttal does not directly answer my original technical questions. The missing hyperparameter specifications, limited sensitivity analysis, incomplete methodological visualization, and unresolved privacy risks remain important concerns.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to ICLR 2027 Reviewers

We sincerely thank all reviewers for their thorough feedback and constructive criticism. We have conducted extensive additional experiments to address the raised concerns, completing 1,140 runs with 5 seeds each across multiple experimental settings. Below we respond to each reviewer's concerns with detailed experimental evidence. All the instructions for the reviewing experiments have been updated to the anonymous open-source link in the document, ensuring the reproducibility and authenticity of the experiments.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.