acceptodds
Under review as a conference paper at ICLR 2027

FedBSE: Federated Bayesian State Evolution for Open-Client Medical Image Segmentation

Abstract

Federated medical image segmentation enables collaborative learning across multiple institutions without sharing raw patient data. However, most existing methods assume a fixed set of participating clients before training, which does not reflect real-world medical collaboration networks where new institutions may continue to join after deployment. Restarting federated training introduces additional communication and optimization costs, while updating the shared model may degrade performance on existing clients. To address this issue, we propose FedBSE, a federated Bayesian state evolution framework for medical image segmentation with open client participation. FedBSE explicitly separates shared segmentation knowledge from client-specific variations and freezes the shared components after the initial federated training stage. For each newly arriving client, the method uses local labeled data and the current state prior to derive a closed-form estimate in a low-dimensional client-state space. This avoids iterative optimization of the full segmentation network and keeps the predictors of existing clients unchanged. The inferred client state and its local posterior uncertainty are then adopted to update the server-maintained state distribution through Gaussian moment matching. The updated distribution serves as the prior for subsequent clients. Experiments on three public multi-center medical image segmentation datasets show that FedBSE achieves competitive segmentation performance while substantially reducing new-client onboarding costs. It also enables the client-state distribution to evolve continuously as new clients arrive.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
federated learning, medical image segmentation, open-client federated learning, Bayesian capability evolution

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2k9QOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FedBSE, a federated segmentation framework that adapts to new hospitals by estimating a low-dimensional client state while keeping the shared model frozen. It then updates a Gaussian state distribution to help subsequent clients, achieving competitive segmentation accuracy with much lower onboarding costs on three medical imaging datasets.

Critical strengths

1 New hospitals can join without bringing existing clients back online or changing their predictors. Separating shared knowledge from client-specific coefficients makes this behavior explicit and preserves existing-client predictions by construction.

2 FedBSE achieves the highest average Dice among the main baselines on all three datasets, with reported local onboarding times of roughly 2–3 seconds. The comparisons with gradient-based coefficient adaptation, diagonal curvature, and full-model fine-tuning help explain where its efficiency comes from.

Critical weaknesses

1 The sequential experiments use only two arrivals per dataset under a preset order, without results across multiple seeds or arrival orders. These examples show that updating the prior can help, but do not establish that it remains reliable over longer sequences or when an unusual client arrives early.

2 The default state has 64 dimensions, but its initial covariance is estimated from only three or five clients, making covariance stabilization especially important. Also, the closed-form solution is exact only for the local quadratic approximation, and the consistency theorem assumes vanishing inference error and uncertainty. The paper does not show when these assumptions hold in practice.

Questions for authors

1 Could you please evaluate more arrival orders and random seeds, including sequences with strongly shifted clients? Does accumulation sometimes hurt later clients?

2 How sensitive are the results to covariance stabilization? With so few initial clients, how does a full covariance compare with diagonal or shrinkage estimates?

3How close is the single closed-form update to a well-converged solution in the same coefficient space, especially under large domain shifts? Could a few additional curvature-based updates improve accuracy at modest cost?

4 In Figure 6, the first PMR client scores 74.99 with the fixed state and 75.20 with the accumulated state, although the text says the two branches are identical for the first arrival. Could you clarify this discrepancy?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2k9Q

Thank you for your constructive comments and valuable suggestions.

Response to Q1&W1: Robustness to Client Arrival Orders

We paired fixed- and updated-prior branches at identical initial checkpoints, data, and arrival orders. Their first-arrival Dice is necessarily identical. Across two initial groups and all orders, the updated prior improves the later client in 4/4 Polyp comparisons and 20/24 PMR comparisons. The four PMR losses all concern C0C_0 after initial group 3,4,5\\{3,4,5\\}; the worst is −1.14-1.14 percentage points (pp). The hard PMR client C0C_0 was tested in every arrival position, so the result is not selected from a single favorable order. Appendix A.6 now reports more order and arriving client. The evidence supports conditional forward transfer and reveals its client-composition boundary, rather than a universal gain.

Table1. Dice of second and third arrivals with accumulated versus fixed priors on Polyp dataset. Cᵢ denotes Site i+1.

Initial clientsArrival orderEvaluated clientDice (Δ pp)
{0,1}2→3C₃71.75% (+1.36)
{0,1}3→2C₂75.69% (+1.80)
{2,3}0→1C₁93.98% (+0.12)
{2,3}1→0C₀71.41% (+0.65)

Table2. Dice of second arrivals with accumulated versus fixed priors on PMR dataset. Cᵢ denotes Site i+1.

Initial clientsArrival orderSecond arrival: Dice (Δ pp)Third arrival: Dice (Δ pp)
{3,4,5}0→1→2C₁ 74.98% (+0.59)C₂ 78.96% (+0.11)
{3,4,5}0→2→1C₂ 78.99% (+0.14)C₁ 74.86% (+0.47)
{3,4,5}1→0→2C₀ 66.85% (-0.13)C₂ 79.65% (+0.80)
{3,4,5}1→2→0C₂ 78.90% (+0.05)C₀ 65.84% (-1.14)
{3,4,5}2→0→1C₀ 66.68% (-0.30)C₁ 76.84% (+2.45)
{3,4,5}2→1→0C₁ 76.24% (+1.85)C₀ 66.80% (-0.18)
{0,1,3}2→4→5C₄ 83.45% (+0.36)C₅ 80.67% (+0.17)
{0,1,3}2→5→4C₅ 81.49% (+0.99)C₄ 83.24% (+0.15)
{0,1,3}4→2→5C₂ 86.39% (+0.03)C₅ 80.70% (+0.20)
{0,1,3}4→5→2C₅ 80.82% (+0.32)C₂ 86.76% (+0.40)
{0,1,3}5→2→4C₂ 86.47% (+0.11)C₄ 83.25% (+0.16)
{0,1,3}5→4→2C₄ 83.59% (+0.50)C₂ 86.56% (+0.20)

Response to Q2&W2: Sensitivity to Covariance Estimation

Sensitivity to Covariance Estimation The observed gains are not contingent on choosing a full covariance estimator. With few initial clients and a 64-dimensional capability vector, estimation is potentially ill-conditioned. We therefore compared full, diagonal, and shrinkage covariance estimators over four Polyp leave-one-client-out splits:

CovarianceDice (%)IoU (%)
Full79.7772.46
Diagonal79.8372.51
Shrinkage79.7272.38

The maximum spread is just 0.11 pp Dice and 0.13 pp IoU, with diagonal covariance performing marginally better than full covariance. These findings suggest that FedBSE exhibits limited empirical sensitivity to the covariance estimation strategy in this setting, even with few initial clients. We have added the protocol and results to Appendix A.7. X\phantom{X} Response to Q3: Closed-Form Inference and Convergence

To assess the accuracy–efficiency trade-off, we compare the single-step closed-form inference with 2 and 5 additional curvature-based refinement steps, as well as well-converged coefficient optimization in the same capability space. The four-fold averaged results on Polyp are reported below.

MethodDice (%)IoU (%)Time (s)
One-step closed-form79.7772.462.67
+2 curvature updates80.3173.0328.22
+5 curvature updates80.4373.1246.95
Well-converged optimization80.4573.13148.42

The single-step closed-form inference achieves performance close to the well-converged solution, with differences of only 0.68 and 0.67 percentage points (pp) in Dice and IoU, respectively, while reducing computation time from 148.42 s to 2.67 s. Additional curvature refinement further narrows this gap. These results indicate that single-step inference provides a favorable accuracy–efficiency trade-off, while a few additional curvature updates can further improve accuracy with diminishing returns. We therefore retain single-step inference as the default for efficient client onboarding. We report the accuracy–cost curve in Appendix A.8. X\phantom{X} Response to Q4: Clarification of Figure 6(b)

We apologize for the unclear presentation in Figure 6(b). It did not show the first client clearly. The corrected figure now shows all three PMR arrivals after initial clients C0,C1,C2\\{C_0,C_1,C_2\\}. The first, C3C_3, is 72.3672.36\\% Dice in both branches. The 74.9974.99\\% fixed and 75.2075.20\\% accumulated bars belong to the second arrival, C4C_4; C5C_5 is third. The caption now names the initial group and order. This makes the paired identity check visible. We have revised the figure caption and axis labels to explicitly indicate the initial clients and the displayed arrival order.

Reviewer_fif9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes FedBSE, a federated Bayesian state evolution framework for medical image segmentation with open client participation. FedBSE parameterizes each client’s model as shared logits plus a client-specific additive correction. For each newly arriving client, FedBSE uses local labeled data and the current state prior to derive a closed-form estimate in a low-dimensional client-state space. Experiments on three multi-site datasets (Polyp, RIF, PMR) under leave-one-client-out show slightly higher average Dice than six baselines, with an onboarding time of about 2.5 s.

Critical strengths
  • This paper targets one solid question for clinical federations in real practice.
  • The method is simple, cheap, and computation-light.
  • The ablations cover the main design choices, and the authors report that full fine-tuning and ReTrain achieve higher new-client accuracy
Critical weaknesses
  • Limited novelty, and missing closely related work. Freezing a shared model and adapting only a low-dimensional client embedding for novel clients is widely used in the works like pFedHN. Shared-representation plus personal-head approaches (FedPer, FedRep, FedBABU) likewise adapt only a small client-specific component. The author need to show the novelty in the key settings.
  • The Bayesian accumulation is conceptually questionable, and the theory is largely far away from real clinical pratice. Σ_t is defined as the population (between-client) covariance, yet the update adds V_n, the posterior uncertainty of one client’s estimate. This mixes estimation uncertainty into population spread. +Inference ablation results need explanation. The “first-order inference” update with beta = 10^-3, which is a gradient step scaled by about 1000 if assume Hn=0. Please double confirm this as shown in A.4 Sensitivity Analysis.
  • All results come from a single run, with no standard deviations or confidence intervals.
Questions for authors

Refer to weakness.

3 comments
AuthorsOpenReview ↗

Response to Reviewer fif9

Thank you for your constructive comments and valuable suggestions.

Response to W1: Novelty Relative to Personalized Federated Learning

We have expanded Section 2.2 to discuss FedPer, FedRep, FedBABU, and pFedHN, including pFedHN's ability to adapt to unseen clients. However, it is worth noting that FedBSE differs fundamentally from these methods in how its decomposition supports subsequent inference and knowledge accumulation. The essential difference lies in what the decomposition enables. Existing approaches personalize network layers, prediction heads, or model parameters generated from client embeddings. FedBSE instead represents client-specific variations through a shared input-dependent basis and capability coefficients: fc(x)=gθ(x)+Φϕ(x)αcf_c(x)=g_\theta(x)+\Phi_\phi(x)\alpha_c (Eq. (2)). With shared components frozen, the prediction Jacobian satisfies ∂fc/∂αc=Φϕ(x)\partial f_c/\partial\alpha_c=\Phi_\phi(x). Thus, the same capability coefficients can be used directly for personalized prediction, prior-guided closed-form inference, and cross-client probabilistic accumulation, rather than serving only as personalized model parameters.

This distinction is supported by our experiments. Removing the capability space, replacing the input-dependent basis, or removing the capability prior reduces Polyp Dice from 79.77% to 75.03%, 78.06%, and 78.79%, respectively (Table 2). More importantly, CCI uses these shared capability coordinates for curvature-aware inference (Eq. (9)), while BCA incorporates each new client's estimated capability and uncertainty into the prior for future arrivals (Eqs. (10)–(11)). This forward knowledge transfer improves performance in 20/24 subsequent PMR evaluations. Therefore, the contribution of FedBSE lies not merely in a different parameter decomposition, but in a decomposition explicitly designed to connect efficient personalization with continuous knowledge accumulation. X\phantom{X}

Response to W2: Bayesian Accumulation and First-Order Scaling

We thank the reviewer for the careful examination of our Bayesian formulation and numerical stability. However, it is worth noting that incorporating local uncertainty into the capability covariance is conceptually consistent. The key distinction is between client heterogeneity and uncertainty about a single global parameter. Eq. (5) defines the initial empirical covariance across client capabilities. Eq. (11) then updates the capability distribution through Gaussian moment matching, explicitly incorporating the new client's inferred capability and its local uncertainty VnV_n. As already explained in Section 3.5, Σt\Sigma_t represents the evolving distribution of client capabilities, not posterior uncertainty about a single shared parameter. These two definitions are consistent. Importantly, VnV_n is not an arbitrary addition: omitting it treats the new client's estimated capability as exact and fails to preserve the mixture's covariance (Theorem 3). Our existing ablation in Appendix A.4, Table 3 directly supports its usefulness: including VnV_n improves Dice/IoU from 72.65%/63.91% to 72.90%/64.32%. Thus, both the mathematical derivation and experimental evidence support uncertainty-aware accumulation.

Regarding first-order scaling, we agree that β=10−3\beta=10^{-3} deserves careful interpretation. When Hn=0H_n=0, Eq. (9) gives αn∗−μt=−1000Σtgn\alpha_n^*-\mu_t=-1000\Sigma_t g_n, not an unscaled −1000gn-1000g_n update. Its magnitude depends on the covariance and gradient, so β−1\beta^{-1} alone does not determine numerical stability. Moreover, full Gauss–Newton inference achieves 79.77% Dice in 2.67 s, compared with 77.96% in 2.35 s for first-order inference, demonstrating the practical value of incorporating curvature. X\phantom{X} Response to W3: Statistical Reliability

The original study evaluates held-out clients across Polyp, RIF, and PMR; controlled component ablations test mechanism dependence. Our extension further covers two initial groups and multiple arrival orders per dataset: 4/4 positive subsequent-client comparisons on Polyp and 20/24 on PMR, with all four PMR failures explicitly reported. This addresses the concern that the accumulation result is specific to a cherry-picked onboarding sequence. Due to the limited rebuttal period, we are unable to complete multi-seed evaluations of the main experiments within this timeframe. We will update the main manuscript once these additional results are available.

Reviewer_fif9OpenReview ↗

follow-up questions

Thank you for the detailed and helpful responses. There are some follow-up questions as shown bellow:

  • "Table 3 shows Vn improves Dice/IoU from 72.65% to 72.90%". Given that the improvement is about 0.25 and the patient number is relatively limited, the author need to show the statistical reliability or confidence for this improvement.
  • "FedBSE instead represents client-specific variations through a shared input-dependent basis and capability coefficients:" it would be great if the author can directly show the effectiveness of this decomposition method under controlled comparison.
  • I agree that updating magnitude depends on the covariance and gradient at the same time in this method, but the performance comparison alone does not establish whether this baseline is appropriately scaled. It would be useful if the author can directly provide the distribution for the covariance and gradient during the training.
AuthorsOpenReview ↗

Response to Reviewer fif9

We sincerely thank the reviewer for the thoughtful follow-up questions and constructive suggestions. We appreciate the opportunity to further clarify our method and strengthen the experimental evidence.

Response to Q1: Statistical Reliability of Uncertainty-Aware Accumulation

To further evaluate the contribution of the local uncertainty term VnV_n, we conducted three independent runs comparing full moment matching with the variant excluding VnV_n under the same evaluation protocol.

Runw/o VnV_n Dice / IoU (%)Full Dice / IoU (%)
172.65 / 63.9272.90 / 64.32
272.58 / 63.8472.90 / 64.32
372.64 / 63.9172.88 / 64.30

The full method consistently outperforms the variant without VnV_n across all three runs, with average improvements of 0.270 percentage points in Dice and 0.425 percentage points in IoU. A two-sided paired t-test on the unrounded results yields p = 0.0100 for Dice and p = 0.0042 for IoU.

These results provide additional evidence supporting the contribution of uncertainty-aware accumulation. Due to the limited rebuttal period, we can currently provide repeated-run statistics only for this specific comparison. We will further report confidence intervals for the main experimental results in the subsequent version of our paper. X\phantom{X}

Response to Q2: Effectiveness of the Capability Decomposition

Our current ablation study in Table 2 provides controlled comparisons that isolate the contributions of the proposed capability decomposition and its input-dependent variation basis. Specifically, replacing the input-dependent variation basis Φϕ(x)\Phi_\phi(x) with an input-independent shared basis reduces Dice from 79.77% to 78.06% and IoU from 72.46% to 69.61%, corresponding to decreases of 1.71 and 2.85 percentage points, respectively. Furthermore, removing the entire client-specific capability term Φϕ(x)αc\Phi_\phi(x)\alpha_c decreases Dice to 75.03% and IoU to 66.95%, corresponding to decreases of 4.74 and 5.51 percentage points, respectively. These controlled comparisons demonstrate that the input-dependent variation basis contributes to segmentation performance beyond merely introducing client-specific capability coefficients, while also supporting the effectiveness of the overall capability representation.

VariantDice (%) ↑IoU (%) ↑
w/o Capability Space75.0366.95
w/o Input-dependent Basis78.0669.61
Full FedBSE79.7772.46

X\phantom{X}

Response to Q3: Numerical Scaling of Capability Inference

To provide more direct numerical evidence, we additionally examined the covariance spectra, gradient norms, and capability update norms across four held-out clients.

Covariance distribution. We have supplemented the eigenvalue spectra of the covariance matrices in Appendix A.10 (Figure 7). The covariance trace, measured by tr(Σt)\mathrm{tr}(\Sigma_t), is 0.00159, 0.00151, 0.00667, and 0.0248 across the four held-out settings, respectively. The spectra show that variance is concentrated in a few principal directions, while the remaining eigenvalues are close to the regularization floor.

Gradient and update magnitudes. The following table reports the observed gradient norms and update-direction norms during capability inference.

Client∥gn∥2\|g_n\|_2∥Δαn∥2\|\Delta\alpha_n\|_2
Leave 02.257910.05920
Leave 10.023120.06303
Leave 20.621520.10151
Leave 30.576440.17559

Although the gradient norms vary substantially across clients, the observed coefficient displacement norms remain within 0.0592–0.1756, without numerical divergence in these evaluations. Together with the covariance spectra, these results provide direct evidence of the numerical behavior during new-client inference.

X\phantom{X}

Thanks again for the insightful comments, which are quite helpful to improve the quality of our paper. We sincerely hope that the Reviewer could kindly reconsider the overall assessment of our work.

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

This paper proposes FedBSE for new clients into a deployed FL medical image segmentation system. The framework separates shared knowledge from client-specific variations through an input-dependent basis and low-dimensional client coefficients. After initial federated training, the shared components are frozen, and new-client coefficients are estimated by solving a locally approximated quadratic objective. Experiments on three medical segmentation datasets demonstrate some segmentation performance improvements.

Critical strengths
  • The paper addresses a practical problem. New institutions may join a medical federated learning system after its initial training has finished. Supporting these institutions without requiring all existing clients to participate in another training process is a relevant and useful objective.
  • The method shows some reasonable performance improvements.
  • The method is easy to understand, and the ablation studies cover the main design.
Critical weaknesses
  • The requirement for labeled data at each new institution limits the applicability. Newly joining hospitals may have little or no annotated data. It remains unclear how performance changes with the annotation budget and whether pseudo-labels generated by the deployed federated model could support adaptation.
  • Existing hospitals could retain their original models while a new hospital fine-tunes its own copy, thereby avoiding changes to existing predictions. Table 2 already shows that full-model fine-tuning achieves higher Dice than FedBSE on Polyp. The presented time cost is not the major bottleneck in real-world.
  • The accumulated knowledge benefits future clients but does not improve existing clients.
  • A single Gaussian may inadequately represent heterogeneous client conditions. Differences in devices, acquisition protocols, or institutional characteristics may produce multiple modes in the capability space.
  • The theorem assumes independently and identically distributed client capabilities, vanishing average squared inference error, and vanishing average local uncertainty. The paper does not establish that its one-step local approximation satisfies these conditions.
  • Several accuracy improvements are modest, and their statistical reliability is unclear.
  • The benefits of historical capability accumulation are not sufficiently justified. Longer sequences, different arrival orders, and shifts in the incoming client population are needed to determine when accumulation helps.
Questions for authors

Please see the weakness part.

2 comments
AuthorsOpenReview ↗

Response to Reviewer iezT

Thank you for your constructive comments and valuable suggestions.

Response to W1: Limited Annotation Budgets

To evaluate FedBSE's performance and adaptation efficiency under limited labeled data, we conducted additional controlled experiments with varying labeled-support budgets on Polyp. The results show that FedBSE achieves higher segmentation accuracy and substantially faster adaptation than independent full-model fine-tuning at both 5% and 10% labeled support. Moreover, FedBSE exhibits a smaller performance degradation as the labeled-support budget decreases. At full support, fine-tuning achieves higher accuracy but incurs substantially greater adaptation cost. Detailed results are provided in Appendix A.9. Zero-label onboarding remains unexplored.

Label budgetFedBSE Dice (%)FT Dice (%)FedBSE time (s)FT time (s)
5%78.5378.260.158.28
10%78.9478.610.2616.37
25%79.5479.870.6636.78
100%79.7781.552.67126.18

X\phantom{X}

Response to W2: Why Not Independently Fine-Tune a Model Copy?

Unlike independent FT, FedBSE not only adapts new clients but also accumulates reusable knowledge for future arrivals. We agree that independent FT can preserve existing predictions. However, it does not inherently transfer newly acquired knowledge to subsequent clients. In contrast, FedBSE continuously updates the capability prior, enabling future clients to benefit from historical information while maintaining competitive segmentation performance and substantially lower adaptation costs, particularly under limited annotations.

Compared with independent FT, FedBSE achieves substantially higher computational efficiency during new-client adaptation. In real-world medical federations, computational efficiency also has meaningful clinical value. Institutions may join continuously, requiring repeated adaptation, evaluation, and deployment. Faster adaptation reduces cumulative computational overhead, facilitates timely model updates, and improves operational flexibility, especially for institutions with limited resources. Although clinical validation remains essential, computational efficiency directly contributes to the scalability of such systems. FedBSE therefore provides a practical balance between segmentation accuracy, efficient onboarding, and continuous knowledge accumulation. X\phantom{X} Response to W3: Existing-Client Performance

Forward accumulation without retrospective model changes is the deployment objective, not an overlooked optimization step. Each new client updates only the capability prior used for future arrivals, leaving shared parameters and previously assigned client coefficients unchanged. Hence previously deployed prediction functions are preserved, avoiding unintended backward interference and changes that might trigger additional site-specific validation. We do not claim retrospective accuracy improvements; the framework deliberately prioritizes stable existing predictions and better-informed future onboarding. X\phantom{X} Response to W4: Adequacy of a Single Gaussian

We use a single Gaussian to model client heterogeneity efficiently, not because we assume all hospitals have similar distributions. Its covariance Σt\Sigma_t captures differences among client capabilities, while CCI uses each new hospital's local data to infer its specific capability rather than simply assigning the global mean. Moreover, with only a few initial clients, estimating multiple Gaussian components could introduce unreliable parameters. Our approach therefore balances heterogeneity modeling and estimation reliability. We acknowledge that a single Gaussian cannot fully represent strongly multimodal distributions, but it provides a practical and efficient approximation for continuous client expansion.

Response to W5: Assumptions of Theorem 4 and One-Step Quality

The useful finite-step guarantees do not depend on the asymptotic assumptions required for the conditional consistency result in Theorem 4. Theorem 3 gives exact first-two-moment matching of the specified accumulated mixture for any finite arrival sequence. Equation (9) exactly solves the local quadratic surrogate, not necessarily the original nonlinear risk. To quantify how much this approximation matters in practice, we compare one step with refinement and well-converged coefficient optimization across four held-out Polyp clients:

MethodDice (%)IoU (%)Time (s)
One-step closed-form79.7772.462.67
+2 curvature updates80.3173.0328.22
+5 curvature updates80.4373.1246.95
Well-converged optimization80.4573.13148.42

The default is just 0.68 pp Dice / 0.67 pp IoU below the converged result while using 55.6× less computation; after five refinements the observed metric gap is nearly closed. This directly evaluates finite-step approximation quality. X\phantom{X} Response to W6: Reliability of Modest Gains

We evaluate FedBSE on three datasets, conduct component ablations, and test multiple client arrival orders. Accumulation improves performance in 4/4 Polyp and 20/24 PMR subsequent-client comparisons, with all four negative cases reported. These results support the effectiveness of forward knowledge accumulation across the tested settings, rather than relying on one favorable result. Due to the limited rebuttal period, we cannot complete additional multi-seed experiments. We acknowledge this limitation and do not claim statistical significance for small differences. We will report the additional results in the revised manuscript once available. X\phantom{X} Response to W7: When Does Historical Accumulation Help?

We paired fixed- and updated-prior branches at identical initial checkpoints, data, and arrival orders. Their first-arrival Dice is necessarily identical. Across two initial groups and all orders, the updated prior improves the later client in 4/4 Polyp comparisons and 20/24 PMR comparisons. The four PMR losses all concern C0C_0 after initial group 3,4,5\\{3,4,5\\}; the worst is −1.14-1.14 percentage points (pp). The hard PMR client C0C_0 was tested in every arrival position, so the result is not selected from a single favorable order. Appendix A.6 now reports more order and arriving client. The evidence supports conditional forward transfer and reveals its client-composition boundary, rather than a universal gain.

Table1. Dice of second and third arrivals with accumulated versus fixed priors on Polyp dataset. Cᵢ denotes Site i+1.

Initial clientsArrival orderEvaluated clientDice (Δ pp)
{0,1}2→3C₃71.75% (+1.36)
{0,1}3→2C₂75.69% (+1.80)
{2,3}0→1C₁93.98% (+0.12)
{2,3}1→0C₀71.41% (+0.65)

Table2. Dice of second arrivals with accumulated versus fixed priors on PMR dataset. Cᵢ denotes Site i+1.

Initial clientsArrival orderSecond arrival: Dice (Δ pp)Third arrival: Dice (Δ pp)
{3,4,5}0→1→2C₁ 74.98% (+0.59)C₂ 78.96% (+0.11)
{3,4,5}0→2→1C₂ 78.99% (+0.14)C₁ 74.86% (+0.47)
{3,4,5}1→0→2C₀ 66.85% (-0.13)C₂ 79.65% (+0.80)
{3,4,5}1→2→0C₂ 78.90% (+0.05)C₀ 65.84% (-1.14)
{3,4,5}2→0→1C₀ 66.68% (-0.30)C₁ 76.84% (+2.45)
{3,4,5}2→1→0C₁ 76.24% (+1.85)C₀ 66.80% (-0.18)
{0,1,3}2→4→5C₄ 83.45% (+0.36)C₅ 80.67% (+0.17)
{0,1,3}2→5→4C₅ 81.49% (+0.99)C₄ 83.24% (+0.15)
{0,1,3}4→2→5C₂ 86.39% (+0.03)C₅ 80.70% (+0.20)
{0,1,3}4→5→2C₅ 80.82% (+0.32)C₂ 86.76% (+0.40)
{0,1,3}5→2→4C₂ 86.47% (+0.11)C₄ 83.25% (+0.16)
{0,1,3}5→4→2C₄ 83.59% (+0.50)C₂ 86.56% (+0.20)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.