acceptodds
Under review as a conference paper at ICLR 2027

Freeing Clients while Enhancing Large Models via Federated Weak-to-Strong Learning

Abstract

Federated fine-tuning enables large models (LMs) to adapt to decentralized private data without raw data sharing. Existing parameter-efficient methods reduce the number of trainable parameters but still require clients to execute large model backbones, imposing substantial memory and computational overhead on resource-constrained devices. To overcome this limitation, we propose FedW2S, a federated weak-to-strong framework that frees up client computational bottlenecks and enhances large models. Instead of deploying a strong model locally, FedW2S transfers the knowledge learned from private data by lightweight weak models on the client side to a strong model on the server side. To mitigate unreliable supervision from weak models, FedW2S combines reliability-based filtering with discrepancy-aware target mixing, selectively absorbing informative weak knowledge while preserving useful knowledge encoded in the strong model. Compatibility-aware strong-to-weak feedback further enables bidirectional knowledge transfer, allowing server-side improvements to propagate back to weak clients and enhance their subsequent supervision. We also conduct a theoretical analysis of this. Experiments across visual and language tasks show that FedW2S achieves competitive strong model performance while significantly reducing the memory and computational requirements for client-side training. The code is available at: https://anonymous.4open.science/r/FedW2S.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, Federated Fine-Tuning, Large Models, Weak-to-Strong Generalization, Parameter-Efficient Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_9QZdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper targets a real bottleneck of federated fine-tuning: parameter-efficient methods reduce trainable parameters but clients still deploy and run the full pretrained backbone—the dominant cost on resource-constrained devices. The authors propose FedW2S, a federated weak-to-strong framework with lightweight weak models on clients and a pretrained strong model permanently on the server. Knowledge transfer uses (i) reliability-based filtering by confidence + top-2 margin and an entropy-driven selection budget; (ii) discrepancy-aware target mixing with a β coefficient from JS divergence and cached-server entropy; (iii) compatibility-aware strong-to-weak feedback gated by confidence margin × JS-based compatibility. The plug-in form L_k = L_base + λ L_S2W works on any base FL. On ImageNet / iWildCam / FMoW (DINO ViT-B/8 → ViT-L/14) and UltraFeedback (Qwen2.5-3B → 7B), FedW2S adds +1.5–2 pp on weak clients and +6–9 pp on strong server over 13 baselines.

Critical strengths
  1. Conceptual clarity about the actual cost of federated LM adaptation. Separating trainable-parameter count from deployed-backbone forward/backward cost and proposing weak-to-strong as a way to address the second (which federated PEFT admittedly does not) is a well-articulated contribution. The headline numbers — ~0.19 GB / 0.01 TFLOPs on the client for ResNet-18 → ViT-L/14, with 87.15 % strong-side on ImageNet and 71.95 % on iWildCam — realise that promise.

  2. Methodologically tight and modular. Each step of FedW2S is a small, auditable component that fits an explicit role: reliability filtering controls which public samples contribute, discrepancy-aware mixing controls how much of the weak teacher to absorb, and the compatibility-aware feedback gate controls when a server prediction is allowed to update the client. The decomposition is strictly additive (Eq. 11) and Table 1 demonstrates consistent +1.5–2 pp on the weak side and +6–9 pp on the strong side across six base FL methods. Ablation (Table 5) shows reliability filtering is the largest single component, target mixing is second, feedback is third.

Critical weaknesses
  1. The end-to-end client-resource advantage is not established. Table 4 reports memory and FLOPs "per input" based on parameter estimation, but FedW2S also requires clients to predict on the public split and, after warm-up, to optimise a feedback loss over that split. In the vision experiments the public split is 40 % of the full dataset while private data are divided among 20 clients; this additional repeated public-data workload can materially change total client FLOPs, wall-clock, energy and storage / I-O. Table 9 similarly reports payload size but not the cost of generating the transmitted predictions or the cumulative communication over the full training run. Measured peak memory, client-hours / FLOPs per round and to convergence, wall-clock, energy if available, and total transmitted bytes are needed under the exact participation and public-data schedules. Without these results, "freeing clients" is demonstrated at the model-size / per-example level but not at the system level.

  2. Performance attribution and statistical support are incomplete, especially for the small margins over strong baselines. Main vision tables provide point estimates averaged over three runs without standard deviations or confidence intervals, even though several claimed advantages are only ~0.1–0.6 percentage points. The language results have strongly overlapping uncertainty ranges. It is also unclear which Table 2 baselines have access to the same unlabelled public inputs and comparable server-side computation; a matched control should adapt the strong model on the same public set without client knowledge, and all relevant baselines should be compared under clearly reported data and compute budgets. The evidence supports "competitive performance," but not reliable superiority over the closest methods.

  3. Heavy dependence on the unlabelled public dataset D_p with limited diagnosis of its distribution shift. The conclusion explicitly states that a large distribution shift between D_p and the client distributions degrades the method, and that weak-to-strong only works when the strong model has enough prior to absorb the weak supervision. Both failure modes matter in real deployments. The paper does not include a controlled shift-experiment (e.g. ImageNet as D_p for iWildCam / FMoW, general web text as D_p for UltraFeedback), nor the per-round reliability survival rate as D_p drifts. Without these, the failure mode is unquantified.

  4. The cost-comparison baseline is not matched, and the trust in Table 2 rests on cost asymmetry. FedW2S's strong-side win (87.15 % on ImageNet vs FedIT 81.04 / FedSLM 79.24 / HSplitLoRA 76.63) is concurrent with the fact that the baselines deploy the strong backbone on the client and FedW2S does not. Table 3 reports the gap (~±15 GB / ~14 TFLOPs on the language task) but does not plot a client memory vs. strong-side accuracy Pareto frontier across methods, leaving open whether FedW2S is on the cost / accuracy frontier or simply the cost-asymmetric best.

  5. The theoretical results are conditional identities / bounds with unknown target risks. Theorem 1 and Proposition 1 are conditional identities / bounds involving unknown target risks, fitting error and unselected-sample drift; they regulate the interpretation of the update but do not by themselves guarantee that FedW2S improves the learned model. The sufficient condition in Theorem 1 is not empirically verified in the evaluated runs, and the framework is per-round with no convergence / stability analysis of the bidirectional feedback loop.

Questions for authors
  1. What are the measured end-to-end client costs per round and to convergence — including local private-data training, all public-set forward passes, feedback training, storage / I-O, and communication — for FedW2S and each resource-oriented baseline? How many public examples or minibatches does each participating client process in every stage?

  2. In Tables 2 and 3, which baselines have access to exactly the same unlabeled public inputs and server-side compute budget? Could you add a matched strong-model control that uses the same public set and cached / self-distillation objective but no client predictions, as well as a compute / data-matched comparison?

  3. Could you report per-run results, standard deviations or confidence intervals, and paired statistical comparisons for Tables 1, 2, 5, 6 and 7? Do the apparent 0.1–0.6 pp gains over the strongest baselines persist across seeds?

  4. How robust is the reliability mechanism when the public distribution is shifted or when clients share a confidently wrong consensus? Can you also report the empirical terms in Theorem 1, particularly the selected-target gain, fitting error and unselected-sample drift, to show whether its sufficient condition is non-vacuous in the evaluated runs?

  5. How sensitive is FedW2S to a distribution-shifted public dataset D_p? Please add a small ablation (or a new table) where D_p is drawn from a different domain than the client task (e.g. ImageNet as D_p for iWildCam / FMoW, general web-crawl text as D_p for UltraFeedback), together with the per-round reliability-filter survival rate.

  6. Could you plot a client memory vs. strong-side accuracy Pareto frontier across FedW2S, FedIT, FedSA-LoRA, FlexLoRA, FLoRG, Fed-PLoRA, HSplitLoRA and FedSLM, both axes, language and vision? This would let the reader see whether FedW2S is genuinely on the cost / accuracy frontier or simply the cost-asymmetric best.

  7. The compatibility gate a_{k,i} combines confidence margin (Eq. 8) and a JS-based compatibility factor (Eq. 9). Could you ablate the two individually and report the correlation between a_{k,i} and true usefulness (e.g. against a small held-out ground-truth public sample)?

9 comments
AuthorsOpenReview ↗

Response to Reviewer 9QZd (1/3)

Thank you for the helpful review. We carefully address your questions and comments below.

Q1&W1 End-to-End Client Resource Costs

We have additionally evaluated end-to-end client resource consumption,including peak GPU memory, training time, average computation, and communication costs. The measured results are presented in Table R2 (Item 2 of the General Response above). Overall, FedW2S maintains a favorable resource-performance trade-off, with peak GPU memory and client computation reduced by 11.63 GB and 2,312 TFLOPs, respectively, compared to LORA-based fine-tuned FL methods.

We would also like to clarify the actual public data usage. The 40% public-data ratio represents the size of the available public data pool, rather than the amount processed by each client in every round. During feedback training, FedW2S employs mini-batch sampling, with only half of each training batch drawn from public data and the remaining half from private data, rather than repeatedly optimizing over the entire public dataset. Furthermore, the feedback phase accounts for only a small portion of the entire federated training process, so it does not impose a significant computational burden on the client.

Q2&W2 Public-Data Access and Matched Comparisons

We clarify that our evaluations already include several public-data-based baselines. FedMD and FedDF (Table 1) utilize public data for knowledge distillation, while FedSLM (Table 2–3) also leverages public data for server-side knowledge transfer. Compared with FedSLM, FedW2S achieves higher average accuracy on ImageNet, iWildCam, and FMoW (81.69/67.49/42.71 vs. 80.86/66.30/42.05, Table 2), as well as competitive language performance (48.5 vs. 47.7, Table 3).

Furthermore, Table R4 below (Appendix B.4 Table 10) compares FedW2S with FedMKT, which requires labeled public data for knowledge selection. Specifically, the two approaches have similar setups, but FedMKT requires the labels in public datasets to enable efficient sample selection for knowledge transfer. As shown in the table R4, FedW2S achieves comparable performance (48.5 vs. 49.1 on average) without using public labels, demonstrating the effectiveness of unlabeled knowledge transfer. While the labeling of this public data provides certain advantages, it also makes it more difficult to adapt the method to real-world scenarios.

Table R4: Strong-model performance comparison between FedW2S and FedMKT on language benchmarks. | Method | AlpacaEval 2 (LC) | AlpacaEval 2 (WR) | Arena-Hard (SC) | Arena-Hard (WR) | Avg. | |:-------|:-----------------:|:-----------------:|:---------------:|:---------------:|:----:| | Base(Strong) | 31.5 ± 0.4 | 30.6 ± 1.3 | 37.3 ± 1.8 | 39.4 ± 1.3 | 34.7 | | FedMKT | 40.5 ± 0.8 | 47.1 ± 1.3 | 50.2 ± 2.4 | 58.7 ± 1.7 | 49.1 | | FedW2S | 39.5 ± 0.3 | 47.3 ± 1.6 | 49.6 ± 1.8 | 57.5 ± 1.5 | 48.5 |

[1] FedSLM. Federated foundation models fine-tuning with heterogeneous compressed clients. 2026. [2] FedMKT. Fedmkt: Federated mutual knowledge transfer for large and small language models. ICCL 2025.

Q3 Statistical Reliability and Small Performance Margins

We thank the reviewer for this valuable suggestion.As stated in Section 5.1, all main results are averaged over three independent runs. To provide a more transparent assessment of performance variability, we have supplemented Appendix B.5 (Tables 11–14) with the standard deviations corresponding to original Tables 1, 2, 5, 6, and 7. Here, we show only the standard deviations for original Table 2 (Table R5). We additionally provide seed-wise paired comparisons against the strongest baselines, as summarized in Table R6 below. These additional results allow us to assess the consistency of performance differences across random seeds.

Overall, across different random seeds, FedW2S generally achieves performance comparable to that of other methods. Importantly, the primary contribution of our work does not hinge on marginal accuracy improvements over the strongest baseline. Instead, FedW2S aims to achieve performance competitive with strong-model-based approaches without requiring the deployment of strong models on resource-constrained clients, thereby substantially reducing client-side memory usage and computational overhead.

AuthorsOpenReview ↗

Response to Reviewer 9QZd (2/3)

Following up on the previous response Q3

Table R5: Standard deviations of strong model accuracy across vision tasks (original Table 2). | Method | ImageNet (ϑ=0.1\vartheta=0.1) | ImageNet (ϑ=1.0\vartheta=1.0) | iWildCam (ϑ=0.1\vartheta=0.1) | iWildCam (ϑ=1.0\vartheta=1.0) | FMoW (ϑ=0.1\vartheta=0.1) | FMoW (ϑ=1.0\vartheta=1.0) | |:---|:---:|:---:|:---:|:---:|:---:|:---:| | Local-Strong | ±2.1\pm 2.1 | ±2.1\pm 2.1 | ±1.7\pm 1.7 | ±2.8\pm 2.8 | ±1.6\pm 1.6 | ±1.8\pm 1.8 | | FedIT | ±2.2\pm 2.2 | ±1.8\pm 1.8 | ±1.0\pm 1.0 | ±2.0\pm 2.0 | ±1.9\pm 1.9 | ±1.6\pm 1.6 | | FedSA-LoRA | ±2.8\pm 2.8 | ±2.4\pm 2.4 | ±1.8\pm 1.8 | ±1.8\pm 1.8 | ±2.1\pm 2.1 | ±2.7\pm 2.7 | | FlexLoRA | ±2.3\pm 2.3 | ±1.9\pm 1.9 | ±1.8\pm 1.8 | ±2.2\pm 2.2 | ±1.5\pm 1.5 | ±2.4\pm 2.4 | | FLoRG | ±2.5\pm 2.5 | ±1.8\pm 1.8 | ±1.4\pm 1.4 | ±1.6\pm 1.6 | ±1.8\pm 1.8 | ±2.4\pm 2.4 | | Fed-PLoRA | ±2.6\pm 2.6 | ±2.3\pm 2.3 | ±1.8\pm 1.8 | ±1.7\pm 1.7 | ±1.2\pm 1.2 | ±1.6\pm 1.6 | | HSplitLoRA | ±1.0\pm 1.0 | ±1.1\pm 1.1 | ±1.8\pm 1.8 | ±1.3\pm 1.3 | ±1.2\pm 1.2 | ±1.9\pm 1.9 | | FedSLM | ±2.5\pm 2.5 | ±2.9\pm 2.9 | ±1.5\pm 1.5 | ±1.8\pm 1.8 | ±1.2\pm 1.2 | ±1.4\pm 1.4 | | FedW2S | ±1.9\pm 1.9 | ±1.6\pm 1.6 | ±1.2\pm 1.2 | ±1.7\pm 1.7 | ±0.9\pm 0.9 | ±0.7\pm 0.7 |

Table R6: Seed-wise paired performance differences between FedW2S and representative baselines. | Source | Dataset/setting | Comparison (A − B) | Δ1\Delta_1 | Δ2\Delta_2 | Δ3\Delta_3 | Δ‾\overline{\Delta} | |:------:|:----------------|:-------------------|----:|----:|----:|----:| | Tab.1 | ImageNet (ϑ = 1, Weak) | FedW2S − FedMD | +1.25 | +1.83 | +0.97 | +1.36 | | Tab.2 | ImageNet (Avg.) | FedW2S − FLoRG | +0.34 | -0.17 | -0.47 | -0.10 | | Tab.2 | iWildCam (Avg.) | FedW2S − HSplitLoRA | +0.61 | +0.49 | +0.46 | +0.52 | | Tab.2 | FMoW (Avg.) | FedW2S − FlexLoRA | +0.63 | +0.38 | +0.52 | +0.51 | | Tab.3 | AlpacaEval 2 | FedW2S − HSplitLoRA | -0.74 | -0.12 | -0.37 | -0.41 | | Tab.3 | Arena-Hard | FedW2S − HSplitLoRA | +0.35 | +0.45 | -0.08 | +0.24 |

Q4&W5 Robustness of Reliability Filtering and Empirical Validation of Theorem 1.

The objective of reliability-based filtering is to reduce the influence of unreliable weak supervision, not to identify every incorrect prediction without ground-truth labels. Our existing Figure 9 provides empirical evidence that the proposed reliability score is positively associated with weak-label correctness across the three vision benchmarks.

To examine the practical implications of Theorem 1, we additionally measure the selected-target gain ASA_S, fitting error δS\delta_S, and unselected-sample drift DSˉD_{\bar S} at three training checkpoints on ImageNet, iWildCam, and FMoW. The results are shown in the Table R7. The results show that the average selected-target gain is positive across all three datasets, while the empirical Brier risk decreases at every evaluated checkpoint. These observations provide empirical evidence of beneficial knowledge transfer in the evaluated settings. Confidently incorrect client consensus and severe public-data distribution shifts remain potential failure modes rather than cases covered by unconditional guarantees.

Table R7: Empirical diagnostics of server-side knowledge transfer. | Dataset | AS↑A_S \uparrow | δS↓\delta_S \downarrow | DS‾↓D_{\overline{S}} \downarrow | ΔBrier↓\Delta \mathrm{Brier} \downarrow | |:---|:---:|:---:|:---:|:---:| | ImageNet | 0.2521 | 0.1874 | 0.0513 | -0.6698 | | iWildCam | 0.2967 | 0.3218 | 0.0157 | -0.3233 | | FMoW | 0.1956 | 0.5061 | 0.0592 | -0.1434 |

Q5&W3 Sensitivity to Public-Data Distribution Shift

To further investigate the sensitivity of FedW2S to public-data distribution shifts, we have conducted additional controlled experiments on iWildCam and FMoW, as reported in Table R1 (Item 1 of our General Response). Specifically, for iWildCam, clients hold private data from disjoint camera views, while the shifted public data are drawn from unseen camera views. For FMoW, clients hold private data from disjoint time periods, while the shifted public data are drawn from unseen periods. We maintain consistent private-data partitions, public-data sizes, and training budgets for fair comparisons.

The results in the table show that our method still demonstrates a certain competitive advantage. Our reliability filtering mechanism reduces the selection budget by increasing prediction uncertainty, thereby limiting the amount of potentially unreliable weakly supervised data transmitted to the server. We acknowledge that significant variations in distribution may still reduce knowledge transfer, particularly when weak predictions are confidently deemed incorrect. We discuss this limitation further in the revised manuscript and plan to conduct in-depth research on it in the future.

AuthorsOpenReview ↗

Response to Reviewer 9QZd (3/3)

Q6&W4 Client Memory–Performance Pareto Comparison

We have added client memory versus strong-model performance Pareto plots for both vision and language tasks in Figures 4 and 10 of this paper, covering FedW2S and all seven federated fine-tuning baselines. The results show that FedW2S lies on the Pareto frontier across the evaluated tasks. For visual tasks, FedW2S achieves an accuracy of 78.51% while using only 0.39 GB of peak GPU memory. By leveraging small model deployment, FedW2S reduces peak memory usage by over 90% compared to the best method FLoRG, while its performance is less than 0.3 points lower. For language tasks, FedW2S achieves good performance while requiring less memory than other methods.These results further highlight that FedW2S provides a favorable trade-off between client resource requirements and strong-model performance, rather than merely improving accuracy through increased client computation.

Q7 Ablation of the Compatibility-Aware Feedback Gate

To examine the individual contributions of the two factors, we conduct additional ablations by separately removing the confidence margin Δk,i\Delta_{k,i} and the compatibility factor gk,ig_{k,i} , while keeping all other training settings unchanged. The results of the elimination experiment are shown in the Table R8. Both factors independently improve performance over the no-feedback variant. Combining them achieves the best results, outperforming confidence-only feedback by 2.3/1.1 percentage points and compatibility-only feedback by 0.8/1.7 points on ImageNet/iWildCam, respectively. The confidence margin restricts feedback based on the server's relative confidence advantage, while the compatibility factor attenuates potentially conflicting guidance. Their combination enables selective absorption of server knowledge while mitigating interference with client-specific learning under non-IID distributions.

Table R8: Ablation study of confidence margin and compatibility factors. | Variant | Weight | ImageNet | iWildCam | |:---|:---:|:---:|:---:| | w/o Feedback | — | 81.2 | 68.9 | | Confidence Only | Δk,i\Delta_{k,i} | 82.6 | 71.4 | | Compatibility Only | gk,ig_{k,i} | 84.1 | 70.8 | | FedW2S (Full) | Δk,igk,i\Delta_{k,i}g_{k,i} | 84.9 | 72.5 |

Thank you again for your constructive suggestions. Could you let us know whether these additions address your questions and concerns? We would be happy to discuss further and provide more details.

Reviewer_2NaPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FedW2S, which uses small client models trained on private data to improve a larger server model through unlabeled public data. It filters unreliable predictions, mixes client and server predictions, and selectively sends knowledge back to clients. Experiments on image classification and language preference alignment show competitive performance with lower estimated client resource requirements.

Critical strengths
  1. Resource-efficient architecture. FedW2S trains lightweight client models while keeping the large model on the server. Knowledge transfer through unlabeled public data supports heterogeneous architectures without requiring clients to run the large model.

  2. Selective bidirectional transfer. Reliability filtering, target mixing, and compatibility-aware feedback provide controls over knowledge exchange, addressing unreliable client supervision and potentially harmful server feedback.

  3. Broad experimental evaluation. The evaluation spans three vision and two language benchmarks and includes multiple federated baselines, non-IID settings, and heterogeneous client architectures. Ablation studies further support the contributions of individual components.

Critical weaknesses
  1. Dependence on suitable public data. FedW2S requires shared, task-relevant public inputs for both W2S and S2W transfer. Obtaining data representative of private client domains may be difficult in specialized applications, limiting practical applicability.

  2. Robustness to public-data mismatch remains unclear. Experiments vary public-data quantity without directly testing missing classes or domain mismatch. Whether knowledge from poorly represented private domains can transfer effectively remains uncertain.

  3. Global selection may overlook difficult classes. Ranking public samples globally by reliability can repeatedly exclude difficult or minority classes.

  4. Feedback can limit useful corrections. Equations (8)–(9) weaken server feedback when it is less confident or strongly disagrees with a client, potentially blocking correct guidance where correction is most needed.

Questions for authors

Please refer to the Weakness part.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 2NaP

Thank you for the helpful review. We carefully address your questions and comments below.

W1 Dependence on Suitable Public Data

We thank the reviewer for highlighting this practical concern. We acknowledge that access to task-relevant public data may limit the applicability of FedW2S in highly specialized domains. However, FedW2S requires only unlabeled public inputs, which can be collected through retrieval and filtering from publicly available data sources without additional annotation costs or access to private client data. These inputs serve as a shared knowledge-transfer interface between heterogeneous weak clients and the strong server model, without requiring clients to deploy the large model backbone. Moreover, the public data need not exactly match each client's private distribution, provided they offer sufficient task-relevant coverage. We conduct additional exploratory experiments in the General Response Table R1. Our reliability-based filtering and discrepancy-aware target mixing further help mitigate unreliable supervision.

W2 Robustness to public-data mismatch remains unclear

We agree that varying public-data quantity alone is insufficient to assess robustness to distribution mismatch. To address this concern, we have conducted additional experiments on iWildCam and FMoW, where the public data are drawn from unseen camera views and time periods, respectively, to simulate public-private domain shifts. The results are presented in Table R1 (Item 2 of the General Response above) and incorporated into Section 5.2 of the revised manuscript. These results show that FedW2S maintains competitive performance under the evaluated distribution shifts, supporting the effectiveness of its selective knowledge transfer mechanisms.

Nevertheless, we acknowledge that severe domain mismatch or missing classes may limit effective knowledge transfer, as the public data may fail to cover important client-specific knowledge. We have clarified this limitation in our revised paper and identified more extreme variations in distribution as an important direction for future work.

W3 Global selection may overlook difficult classes

Regarding the potential for overlooking difficulties or minority samples, our FedW2S partially mitigates this issue through two complementary design components. Our reliability-based filtering mechanism is designed to improve the reliability of weakly supervised samples.

First, Eq.(4) adopts an entropy-adaptive selection budget, Mp=⌊np(1−Hw)⌋M_p=\lfloor n_p(1- \mathcal{H} _w)\rfloor, rather than a fixed selection size. As the uncertainty of weak predictions decreases, the budget can expand, allowing broader sample coverage in subsequent rounds.

Second, Eq.(8)–(9) deliberately regulate server feedback through ak,i=Δk,igk,ia_{k,i}=\Delta_{k,i}g_{k,i}, jointly considering the server's relative confidence advantage and predictive compatibility. Combined with the private-data objective in Eq.(11), this selective feedback helps prevent server knowledge from overriding valuable client-specific knowledge, including that associated with minority classes, thereby preserving potentially complementary supervision for subsequent weak-to-strong transfer.

W4 Feedback can limit useful corrections

We acknowledge that selective feedback may suppress some beneficial corrections. This reflects a deliberate trade-off between absorbing strong model knowledge and preserving client specific knowledge. In non-IID federated settings, clients may learn valuable domain specific patterns, including those associated with minority classes. Therefore, disagreement with the server does not necessarily imply incorrect client predictions. Unrestricted feedback could overwrite such complementary knowledge and weaken subsequent weak-to-strong transfer.

To mitigate this risk, Eqs. (8)–(9) regulate feedback through ak,i=stopgrad⁡(Δk,igk,i)a_{k,i}=\operatorname{stopgrad}(\Delta_{k,i}g_{k,i}), where Δk,i\Delta_{k,i} captures the server's relative confidence advantage and gk,ig_{k,i} continuously attenuates feedback according to predictive compatibility, rather than imposing a hard agreement constraint. Combined with the private-data objective in Eq. (11), this design limits interference with personalized learning while allowing selective knowledge absorption. The goal is to mitigate the impact of potentially harmful consensus feedback while retaining personalized client knowledge, rather than fully adopting consensus knowledge.

Thank you again for your constructive suggestions. Could you let us know whether these additions address your questions and concerns? We would be happy to discuss further and provide more details.

AuthorsOpenReview ↗

Follow up on our responses

Dear Reviewer 2NaP,

Thank you again for your thoughtful review and constructive feedback. We would like to kindly follow up on our previous response. We have carefully addressed your concerns and provided additional experimental results on public-private distribution shifts (Table R1), along with further clarifications regarding public data requirements, sample selection, and selective feedback mechanisms.

We would greatly appreciate it if you could take a moment to review our response when convenient and let us know whether our additional results and explanations have helped address your concerns. We would be more than happy to discuss any remaining questions or provide further clarification.

Thank you very much for your time and consideration.

Reviewer_J4mKOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes FedW2S, a federated weak-to-strong learning framework that enables resource-constrained clients to contribute to the improvement of a pretrained strong model without locally executing the large model backbone. FedW2S uses lightweight client-side models to transfer task-specific knowledge to the server through three mechanisms: reliability-based filtering, discrepancy-aware target mixing, and compatibility-aware strong-to-weak feedback. The paper also provides theoretical analyses of the server-side risk and client-side feedback. Experiments on vision and language tasks compare FedW2S with existing federated learning and federated fine-tuning methods. The reported results demonstrate competitive model performance while reducing client-side computational and memory costs.

Critical strengths
  1. The paper addresses the computational and memory bottlenecks of federated fine-tuning by keeping lightweight models on clients and a strong model on the server.
  2. The proposed reliability filtering, target mixing, and bidirectional knowledge transfer mechanisms address relevant challenges in weak-to-strong learning.
  3. Experiments cover vision and language tasks, with comparisons against multiple baselines and ablation studies supporting the effectiveness of the proposed approach.
Critical weaknesses
  1. The proposed mechanisms resemble existing techniques in knowledge distillation, confidence-based filtering, and selective knowledge transfer. Their combination is reasonable, but the paper does not sufficiently establish a fundamentally new learning principle.
  2. The theoretical results rely on conditional assumptions and do not establish that the proposed mechanisms necessarily improve generalization or the overall learning process.
  3. The evaluation could more thoroughly examine performance under limited or distribution-shifted public data and provide clearer evidence on run-to-run variability and computational cost measurements.
Questions for authors
  1. What fundamentally distinguishes FedW2S from existing federated distillation and weak-to-strong learning methods?
  2. Why are all three proposed components necessary, and do they offer clear advantages over simpler alternatives?
  3. Under what conditions can the theoretical analysis guarantee beneficial knowledge transfer rather than merely bound potential degradation?
  4. Does FedW2S remain effective with limited or distribution-shifted public data, and are the improvements statistically significant?
  5. Are the reported computation and memory costs measured on actual hardware, and are comparisons with baselines conducted under fair settings?
5 comments
AuthorsOpenReview ↗

Response for Reviewer J4mK (1/2)

Thank you for the helpful review. We carefully address your questions and comments below.

Q1&W1 The fundamental distinctions from existing methods

The key distinction of FedW2S lies in its federated weak-to-strong knowledge transfer framework, which enables resource-constrained clients to collaboratively enhance large cloud-based models. Existing federated distillation methods (such as FedMD and FedDF) primarily focus on knowledge sharing or aggregation among client models, but struggle to address knowledge transfer between models of asymmetric sizes. Traditional weak-to-strong learning primarily focuses on knowledge transfer from weak teachers to strong students, but it struggles to address the challenges of multi-round transfer among heterogeneous clients in federated learning.

To bridge this gap, FedW2S jointly regulates which weak knowledge to transfer (Reliability-Based Filtering), how much to absorb while retaining strong-model knowledge (Discrepancy-Aware Target Mixing), and when to provide feedback without overriding client-specific knowledge (Compatibility-Aware Feedback). These mechanisms form a unified federated learning process. We further clarify these distinctions in Appendix C.

Q2 The necessity and advantages of the three proposed components

The three components address distinct yet complementary challenges in bidirectional weak-to-strong knowledge transfer: (1) Reliability-Based Filtering dynamically identifies reliable client knowledge, mitigating the propagation of noisy supervision that fixed-threshold filtering cannot adequately handle in evolving federated settings; (2) Discrepancy-Aware Target Mixing balances weak supervision with the strong model's existing knowledge, preventing over-reliance on noisy signals and facilitating effective weak-to-strong transfer; and (3) Compatibility-Aware Feedback selectively transfers enhanced knowledge back to clients based on compatibility, improving local performance while minimizing interference with personalization. Together, these components coordinate knowledge reliability, transfer strength, and feedback compatibility rather than merely combining independent distillation techniques.

The experimental results also verify these components. The ablation results in Table 5 show consistent performance degradation when any component is removed, confirming their complementary contributions. Furthermore, Table R3 below (Table 6 in Appendix B.2) compares FedW2S with simpler alternatives. Across six settings, FedW2S achieves average accuracy improvements of 9.04 and 7.14 percentage points over Vanilla KD and Vanilla Bidirectional KD, respectively. It also outperforms the best fixed-threshold filtering baseline by 2.32 percentage points on average. These results demonstrate the clear advantages of the proposed coordinated mechanism over simpler knowledge transfer strategies.

Table R3: Performance comparison with simpler alternatives (%). | Variants | ImageNet ϑ=0.1\vartheta=0.1 | ImageNet ϑ=1\vartheta=1 | iWildCam ϑ=0.1\vartheta=0.1 | iWildCam ϑ=1\vartheta=1 | FMoW ϑ=0.1\vartheta=0.1 | FMoW ϑ=1\vartheta=1 | |:---|---:|---:|---:|---:|---:|---:| | Vanilla KD | 64.58 | 74.31 | 54.92 | 65.23 | 31.76 | 38.75 | | Vanilla bidirectional KD | 67.36 | 76.94 | 55.39 | 66.83 | 32.74 | 41.68 | | Fixed-threshold filtering G(>0.2)\mathcal{G}(> 0.2) | 72.37 | 81.48 | 59.84 | 69.55 | 34.46 | 45.38 | | Fixed-threshold filtering G(>0.4)\mathcal{G}(> 0.4) | 76.87 | 81.61 | 60.37 | 69.89 | 35.74 | 45.26 | | FedW2S (Ours) | 78.51 | 84.87 | 62.48 | 72.50 | 37.85 | 47.56 |

Q3&W2 Conditions for theoretical analysis

Theorem 1 establishes an explicit sufficient condition for beneficial server-side knowledge transfer. Specifically, Eq. (12) shows that the mixed target achieves a lower conditional Brier risk than the original strong-model prediction when bi(wi)−bi(ui)<βi∥ui−wi∥22b_i(w_i)-b_i(u_i)<\beta_i\|u_i-w_i\|_2^2. Importantly, this condition can hold even when the weak teacher has a higher individual risk than the strong model, demonstrating how complementary weak supervision can produce a beneficial mixed target.

Furthermore, Theorem 1 guarantees a strict reduction in the average conditional Brier risk over the fixed public inputs when ρAS>22(ρδS+DSˉ),\rho A_S>2\sqrt{2}\left(\rho\sqrt{\delta_S}+D_{\bar S}\right), where ASA_S denotes the mixed-target gain, δS\delta_S the fitting error, and DSˉD_{\bar S} the unselected-sample prediction drift. Thus, beneficial transfer is guaranteed when the gain from complementary supervision outweighs fitting error and unintended prediction drift. We conduct an additional experiment to verify this condition; please see the response to Question Q4&W5 from reviewer 9QZd.

AuthorsOpenReview ↗

Response for Reviewer J4mK (2/2)

Q4&W3 Robustness to Limited and Distribution-Shifted Public Data

Regarding limited public data, Figure 5(b) already investigates the effect of varying the public-data ratio. The results show that performance generally improves with additional public data, while FedW2S remains effective with reduced public-data availability.

To further address distribution mismatch, we have conducted additional experiments on iWildCam and FMoW, where public data are drawn from unseen camera views and time periods, respectively. The results are presented in Table R1 (Item 1 of our General Response above) and have been incorporated into Table 6 of the revised Section 5.2. FedW2S maintains competitive performance under both distribution shifts, supporting the effectiveness of selective knowledge transfer in these settings.

Q5 Actual hardware performance measurements and fairness

To supplement the comparison of actual hardware resources, we conduct an additional hardware-level resource evaluation on ImageNet (ϑ=0.1\vartheta=0.1), including peak and average GPU memory usage, computation per round, cumulative communication, and execution time. The results are shown in Table R2 (Item 2 in the General Response above). Compared to LORA fine-tuning of large models deployed at the edge, FedW2S reduces the peak client memory and computational requirements per round to 0.39GB and 181.04 TFLOPs, respectively, while maintaining competitive and robust model performance.

Regarding fairness, all methods follow consistent data partitioning, client participation, and evaluation protocols, as detailed in Appendix A. In our main experiment, many baseline methods require clients to deploy large models for fine-tuning. But our method isn't necessary. And the FedDF, FedMD, and FedSLM methods utilize public data strategies on the server side in Table 1-3. Additionally, in Appendix B.4, we compare FedMKT, an algorithm that relies on labeled public datasets. For this result, please also view Reviewer 9QZd’s response to Q2&W2.

Thank you again for your constructive suggestions. Could you let us know whether these additions address your questions and concerns? We would be happy to discuss further and provide more details.

AuthorsOpenReview ↗

Follow up on our responses

Dear Reviewer J4mK,

Thank you again for your thoughtful review and constructive suggestions. We have carefully addressed your concerns by further clarifying the key distinctions of FedW2S from existing methods, providing additional comparisons with simpler alternatives, elaborating on the theoretical conditions for beneficial knowledge transfer, and conducting additional experiments on public-data distribution shifts and hardware resource measurements.

We would sincerely appreciate it if you could take a moment to review our responses when convenient and let us know whether these additional explanations and results have helped address your concerns. We would be more than happy to discuss any remaining questions or provide further clarification.

Thank you very much for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their constructive comments and valuable suggestions. We appreciate their recognition of FedW2S's resource-efficient architecture and selective bidirectional knowledge transfer. To address the common concerns, we provide additional analyses and clarify two key aspects: (1) robustness to public data distribution shifts and (2) actual Client-Side resource efficiency.

Robustness to Public Data Distribution Shifts

We thank the reviewers for pointing out the importance of mismatch in the distribution of public data. To further investigate this issue, we conduct additional experiments under two controlled domain-shift settings. Specifically, for iWildCam, clients hold private data from disjoint camera views, while the shifted public data are drawn from unseen camera views. For FMoW, clients hold private data from disjoint time periods, while the shifted public data are drawn from unseen periods. All methods are evaluated under consistent experimental settings to ensure fair comparisons.

As shown in Table R1, FedW2S maintains competitive performance under both distribution shifts, achieving 68.94% on iWildCam and 42.57% on FMoW. Specifically, FedW2S outperforms the strongest baseline by 0.62 percentage points on iWildCam, while achieving performance within 0.11 percentage points of the best baseline on FMoW. These results suggest that FedW2S can maintain effective knowledge transfer under the evaluated public-private distribution shifts, supporting the robustness of its reliability-based filtering and discrepancy-aware target mixing mechanisms. We have included these additional experiments in Section 5.2.

Table R1: Classification accuracy under public–private distribution shifts on iWildCam and FMoW (%). |Dataset|FedAvg|FedIT|FedSA-LoRA|FlexLoRA|FLoRG|Fed-PLoRA|HSplitLoRA|FedSLM|FedW2S| |---|---|---|---|---|---|---|---|---|---| |iWildCam|62.67|65.72|66.39|66.61|67.27|67.84|67.35|68.32|68.94| |FMoW |37.08|39.57|40.13|41.73|40.96|41.42|42.68|41.64|42.57|

Actual Client-Side Resource Efficiency

To provide a more comprehensive assessment of client-side resource efficiency, we conduct additional hardware-level evaluations on ImageNet (ϑ=0.1\vartheta=0.1), using ResNet-18 as the client model and ViT-B/8 as the server model, with a batch size of 16. We measure peak and average GPU memory, average FLOPs per round, average communication, and average execution time, as summarized in Table R2.

The results demonstrate that FedW2S substantially reduces client-side resource requirements compared with strong model baselines. Specifically, compared to deploying a large model using the LoRA fine-tuning method, FedW2S reduces peak GPU memory usage from 12.02 to 0.39 GB, decreases the average computational load per round from 2,493.08 to 181.04 TFLOPs, and shortens the average execution time from approximately 116 to 26.8 seconds. Since FedW2S does not communicate LoRA parameters, it incurs higher communication costs. In language tasks, this communication cost is also similar under the same LoRA-based training strategy.

Table R2: Measured client-Side resource costs across federated learning methods on ImageNet. | Metric | FedAvg | FedMD | FedSLM | FedIT | FlexLoRA | Fed-PLoRA | FLoRG | FedSA-LoRA | HSplitLoRA | FedW2S | |---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:| | Peak GPU Memory (GB) | 0.37 | 0.79 | 10.93 | 12.02 | 12.02 | 12.02 | 12.02 | 12.02 | 6.44 | 0.39 | | Avg. GPU Memory (GB) | 0.28 | 0.36 | 1.43 | 1.67 | 1.62 | 5.72 | 1.72 | 1.58 | 2.72 | 0.33 | | Avg. TFLOPs | 79.37 | 472.79 | 1695.69 | 2493.08 | 2493.08 | 2493.08 | 2493.08 | 2493.08 | 1237.01 | 181.04 | | Com. (MB) | 85.7 | 135.1 | 11.7 | 3.4 | 3.4 | 3.4 | 3.4 | 1.2 | 17167.6 | 110.4 | | Avg. Time (s) | 16.8 | 38.2 | 101.6 | 116.4 | 115.2 | 125.1 | 109.6 | 116.8 | 58.6 | 26.8 |

Summary:

Beyond the robustness and resource evaluations presented above, we further clarify the conditions and limitations of beneficial knowledge transfer, provide standard deviations and seed-wise comparisons to assess statistical reliability, and introduce memory–performance Pareto analyses across vision and language tasks. Additional ablation studies further examine the contributions of the proposed knowledge transfer mechanisms. These results are all included in the “Reviewer Responses” portion.

We sincerely appreciate the reviewers' comments, which has helped us improve the rigor, clarity, and completeness of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.