{"id":"07dae7d2-c1bc-4a77-a122-a676068f89f2","arxiv_id":"2502.09104","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey that organizes one-shot federated learning methods into four categories (parameter learning, knowledge distillation, generative models, ensemble methods) and outlines future research directions.","lead":"This paper is a survey of one-shot federated learning, where clients communicate with a server only once. It groups existing methods into four technique families and outlines future research directions, including applications to large language models.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 4.1 findings rely on unsupported cross-paper comparisons; the practical-guide recommendations are not empirically grounded.","rationale":"The reader's weakest assumption accurately identifies the load-bearing issue: Section 4.1 draws practical conclusions from cross-paper comparisons with non-identical experimental settings. This is the most important weakness because the paper's contribution is framed as a 'practical guide', and the findings are explicitly intended to guide future research. The survey does not report any new experiments, so these findings must rest on the cited papers' evidence, but the cited papers rarely compare methods across categories under controlled conditions. I agree that a conditional verdict requiring a limitation statement or a systematic methodology is appropriate. I considered whether the 'first survey' novelty claim is more load-bearing, but there is no internal evidence contradicting it, and it is not the central assumption on which the paper's usefulness depends. The concrete test I propose would directly challenge the key finding about generative models versus distillation and the adaptive-ensemble advantage; if the test fails, the paper's recommendations would need substantial revision. I therefore recommend keeping the conditional verdict.","tokens_in":13974,"tokens_out":4706,"duration_ms":49502,"concrete_test":"Run a controlled head-to-head comparison on a fixed benchmark: CIFAR-10, ResNet-18, 10 clients with label skew (Dirichlet alpha=0.5), and identical local training budgets. Evaluate one representative method per category: FedKT (distillation), FedDISC or FedDEO (generative, with and without the pre-trained foundation model), FENS (adaptive ensemble), and FedAvg (parameter learning). Record global-model test accuracy. If the generative method's advantage over distillation disappears or reverses when the foundation model is removed, or if the adaptive ensemble does not beat a static ensemble under the same setup, the Section 4.1 findings are not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central value as a practical guide rests on the findings in Section 4.1, especially the claims that generative models outperform knowledge distillation methods and that adaptive ensemble methods yield better results than static ones. These are presented as general conclusions but are derived from informal reading of papers that differ in datasets, model architectures, communication constraints, and evaluation protocols. The survey provides no quantitative synthesis, no unified benchmark, and no controlled comparison. For example, the claim about generative models may be confounded by the use of pre-trained foundation models: FedDISC, FedDEO, and FedBiP leverage CLIP or Latent Diffusion, whereas distillation methods like FedKT or One-shot FL typically assume no such external priors. The observed advantage could stem from the foundation model rather than the OFL technique. Similarly, the adaptive-ensemble claim is based on methods that also introduce other components (e.g., Co-Boosting reinforces via data generation; IntactOFL uses a mixture-of-experts), so the isolated effect of adaptivity is unidentified. Since Section 4.1 concludes with recommendations for future research, unsupported findings directly undermine the paper's stated purpose of providing guidance and insights. The reader's conditional verdict correctly identifies this as the key weakness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a survey of one-shot federated learning (OFL), a paradigm that restricts client-server communication to a single round. It proposes a taxonomy of OFL methods into four groups (parameter learning, knowledge distillation, generative models, and ensemble methods) plus hybrid methods, discusses the two main challenges (data and model heterogeneity), and derives a set of practical findings and future directions. The authors claim this is the first survey specifically focused on OFL and that it comprehensively covers all existing OFL literature.","tokens_in":14181,"tokens_out":5003,"duration_ms":47516,"significance":"If the taxonomy and coverage are accurate, the survey provides a useful structured entry point to a rapidly growing subfield, and the discussion of hybrid methods is a useful contribution. The paper's strengths include a coherent high-level categorization and a synoptic table that helps readers compare methods. However, the practical-guide findings in Section 4.1 rest on informal cross-paper comparisons rather than on any controlled evidence, which substantially limits the reliability of the guidance. The paper also makes strong completeness claims without providing a systematic selection protocol. These issues reduce the significance of the survey as a practical reference unless they are addressed.","major_comments":[{"comment":"The comparative findings in Section 4.1 (e.g., 'Generative models tend to perform better compared to knowledge distillation methods' and 'adaptive ensemble methods typically yield better results') are presented as general conclusions but are derived from informal reading of papers that differ in datasets, model architectures, communication constraints, and evaluation protocols. No quantitative synthesis or unified benchmark is provided. The comparisons are confounded: generative methods such as FedDISC, FedBiP, and FedDEO use pretrained foundation models (CLIP, latent diffusion), whereas distillation methods like FedKT and One-shot FL do not, so the observed advantage may stem from the foundation model rather than the OFL technique. Similarly, adaptive ensemble methods such as IntactOFL and Co-Boosting also introduce data generation or mixture-of-experts, so the isolated effect of adaptivity is unidentified. Because the paper's stated purpose is to 'provide guidance and insights for future research' (Section 1), these findings need to be either supported by a controlled comparison or qualified with explicit caveats about the confounds and the limited comparability of the underlying experiments.","section":"Section 4.1 and Section 3.4"},{"comment":"The paper repeatedly claims to cover 'all current methods of OFL' (Section 3.5), to be 'the first survey paper specifically focused on the OFL domain' (Section 1), and to 'comprehensively cover all existing OFL literature' (Section 5). However, the manuscript does not describe a systematic search protocol, inclusion criteria, or the period of coverage. Without such a methodology, the completeness claims are not verifiable. The authors should either add a methodology subsection describing how papers were identified and selected, or temper the claims to describe a representative overview rather than an exhaustive one.","section":"Sections 1, 3.5, and 5"}],"minor_comments":[{"comment":"The problem formulation in Eq. (1) is not mathematically well-defined: as written, α_i multiplies the entire sum (and α_i is indexed by i but appears outside the sum), and the relationship between the global parameter w and the local parameters w_i is not specified. Please rewrite the objective so that α_i appropriately weights each local loss and state how w is obtained from {w_i}.","section":"Section 2.1, Eq. (1)"},{"comment":"The phrase 'This challenge is essentially an out-of-distribution (OOD) detection problem' is imprecise, since OOD detection usually refers to detecting samples outside the training distribution at inference time, whereas the issue here is about generalization under distribution shift. Consider rephrasing.","section":"Section 2.2"},{"comment":"The description of k-FED is ambiguous about what exactly is clustered (i.e., local model parameters) and what the server does with the cluster means. Please clarify.","section":"Section 3.1"},{"comment":"The text refers to 'FENSE' but the cited method is 'FENS' (Allouah et al., 2024). Please correct the typo.","section":"Section 4.1"},{"comment":"The row for FedDEO does not list 'image captions' under 'Additional Data', even though Section 3.3 states that the method transmits image captions. Please make the table consistent with the text.","section":"Table 1"},{"comment":"The entries '[Liu and Zeng, 2024 ]' and '[Shah and Lau, 2021 ]' contain spaces before the closing bracket; please fix the formatting.","section":"References"},{"comment":"The citation to Alemdar et al. (2021) to support the claim that OFL 'alleviates the requirements for transmission synchronization' seems out of place, as that reference concerns RF clock synchronization in wireless networks rather than OFL. Please either clarify the connection or replace the citation.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The reference list and the highlighted findings include several works from the same group (FuseFL, FedLPA, HPFL), and these methods are often presented as promising. While this does not by itself undermine the taxonomy, the authors should be mindful of a potential self-promotion bias, and the editors may want to verify the 'first survey' claim and the completeness of the coverage against other recent surveys or position papers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a survey, not a research paper, and the main contribution is organizational. The four-category taxonomy (parameter learning, knowledge distillation, generative models, ensemble methods) is mostly sensible, and the synoptic table is genuinely useful. It covers roughly 25 OFL methods, including very recent work, and the descriptions are largely faithful to the underlying papers. The discussion of hybrid methods is helpful, and the authors generally separate description from their own commentary.\n\nThe soft spots are real but not fatal. Section 4.1 presents 'findings' like generative models outperforming knowledge distillation and adaptive ensembles beating static ones. These are informal observations across papers that differ in datasets, architectures, and evaluation protocols. The stress-test note is right: methods like FedDISC, FedDEO, and FedBiP rely on pre-trained foundation models (CLIP, Latent Diffusion), so any performance advantage may come from the foundation model rather than the OFL technique. The adaptive-ensemble claim is similarly confounded because the methods differ in many other ways. As a 'practical guide,' this should have been flagged as a limitation. There is no systematic search protocol, so 'comprehensive' and 'first survey' are overclaims. The problem formulation in Eq. (1) is sloppy: the α_i weight is placed oddly, and the relationship between w and w_i is unclear. Minor, but it sets a careless tone.\n\nThe self-citations (FuseFL, FedLPA, HPFL) are noticeable but the descriptions seem fair, and the taxonomy does not reduce to those works. I do not see this as a self-promotion problem.\n\nWho is this for? A newcomer to OFL who wants a quick map of the field. It is not a rigorous benchmark or meta-analysis. It deserves peer review, because a good referee can push the authors to either temper the findings or add a real methodology section and limitation statement. I would not desk-reject it.","headline":"Useful survey of one-shot FL with a sensible taxonomy; the 'findings' in Section 4.1 are informal observations, not validated comparisons, and the paper overclaims completeness.","tokens_in":14695,"tokens_out":2680,"would_cite":true,"duration_ms":26200,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper surveys one-shot federated learning, groups the methods into four technique families, and draws practical design guidance from cross-method comparisons.","keywords":["one-shot federated learning","federated learning","data heterogeneity","model heterogeneity","knowledge distillation","generative models","ensemble methods","taxonomy"],"falsifier":"Run the surveyed methods on a single fixed benchmark with the same datasets, model architectures, compute budgets, and privacy settings; if generative models do not systematically beat knowledge distillation, or adaptive ensembles do not beat static ones, the paper's central comparative findings fail.","tokens_in":13740,"feed_emoji":"🤝","tokens_out":5333,"duration_ms":48763,"temperature":0.7,"pith_summary":"One-shot federated learning (OFL) cuts client-server communication to a single round. The paper argues that the field has grown enough to need a dedicated map, and that none of the existing surveys covers it. It proposes a taxonomy of OFL techniques into four families (parameter learning, knowledge distillation, generative models, and ensemble methods) plus hybrids, and uses that taxonomy to compare the methods' designs, privacy properties, and accuracy behavior. The practical payoff is a set of findings: generative models tend to outperform distillation, adaptive ensembles outperform static ones, prototype learning is a promising component, and multi-step server-side updates can improve results within the one-round constraint.","feed_headline":"One-shot federated learning gets a practical taxonomy","feed_subtitle":"A survey sorts OFL methods into four families and names the design choices that work best.","key_machinery":"The central object is the taxonomy itself: four main families — Parameter Learning, Knowledge Distillation, Generative Models, and Ensemble Methods — with hybrid methods that combine them. The taxonomy does the work of the argument: it gives each surveyed method a location, exposes which techniques address data heterogeneity versus model heterogeneity, and organizes the cross-method comparisons that support the paper's findings. Within the families, the paper also names recurring mechanisms such as prototype learning, empirical Fisher aggregation, data and model distillation, GAN/VAE/diffusion-based data generation, and static versus adaptive ensembling.","core_discovery":"On its own terms, the paper's central claim is that one-shot federated learning can be systematically organized and that doing so reveals consistent patterns: the field faces data heterogeneity and model heterogeneity, methods fall into four technique families, and several design choices reliably help accuracy. The paper presents itself as the first survey focused specifically on OFL, with a taxonomy that separates methods by the technique they employ rather than by the problem they address. It reports findings such as generative models generally beating knowledge distillation, adaptive ensembles beating static ones, prototype learning being a strong building block, and performing multiple server-side updates after a single upload being beneficial.","pith_inferences":["A natural next step the survey does not take is a unified benchmark that fixes datasets, architectures, and budgets across all four families, which would either confirm or overturn the ranking of generative over distillation and adaptive over static ensembles.","The trend the paper documents, from GANs to VAEs to diffusion models, suggests that OFL accuracy will keep tracking the frontier of generative modeling; a practical extension would be a modular library where the data generator is swappable.","For LLM-scale federated learning, the survey's 'upload once, update many times' finding points to a concrete research program: merging or distilling fine-tuned LLMs in a single round while running iterative server-side refinement without additional client round trips.","The privacy discussion is mostly qualitative; a sharper extension would analyze each taxonomy family under a formal differential-privacy budget and characterize the accuracy-privacy-communication trade-off."],"forward_implications":["If the taxonomy is right, new OFL papers can be classified quickly, and the field can accumulate results by technique family rather than by individual method.","If generative models generally outperform knowledge distillation in one-shot settings, then future OFL work should prioritize data-generation approaches and track the quality of the chosen generator.","If adaptive ensembles reliably beat static ones, then server-side aggregation should be designed to weigh local models per data instance or per similarity, as in mixture-of-experts approaches.","If multiple server-side updates after one upload help, then the one-shot constraint should be understood as restricting client-server communication, not the server's internal optimization.","If prototype learning is a broadly compatible component, then new methods can graft prototype extraction onto distillation-, generative-, or ensemble-based pipelines."],"supporting_citations":[{"why":"Introduces one-shot federated learning and the distillation-plus-ensemble template that anchors the hybrid category.","marker":"[Guha et al., 2019]"},{"why":"Supplies the Bayesian nonparametric prototype/matching approach that parameter learning later extends.","marker":"[Yurochkin et al., 2019]"},{"why":"Extends prototype matching to CNNs and LSTMs, forming the FedMA backbone within parameter learning.","marker":"[Wang et al., 2020]"},{"why":"FedKT is a key distillation-plus-ensemble hybrid with public data, used to compare distillation order.","marker":"[Li et al., 2021b]"},{"why":"Dense is the data-free generative-plus-ensemble-plus-distillation baseline that later hybrid methods measure against.","marker":"[Zhang et al., 2022]"},{"why":"Co-Boosting provides evidence for the finding that iterative server-side updating after one upload improves accuracy.","marker":"[Dai et al., 2024]"},{"why":"FENS provides evidence for the claim that adaptive ensemble weighting beats static ensembling.","marker":"[Allouah et al., 2024]"},{"why":"FuseFL supports both adaptive ensembling and the bottom-up multi-iteration update finding.","marker":"[Tang et al., 2024b]"},{"why":"IntactOFL's mixture-of-experts adaptive ensemble outperforms Dense and Co-Boosting, supporting sample-adaptive ensembling.","marker":"[Zeng et al., 2024]"}],"fun_headline_variants":["One-shot FL: a practical map of methods and tradeoffs","Four families of one-shot federated learning, decoded","One-shot FL survey: generative models beat distillation","A hands-on taxonomy for one-shot federated learning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The findings in Section 4.1 compare accuracy numbers reported in different papers that use different datasets, model architectures, and experimental settings, and the paper provides no unified benchmark to guarantee those comparisons are fair; if the settings are not comparable, the rankings may not hold.","fun_headline_variants_meta":{"raw":{"variants":["One-shot FL: a practical map of methods and tradeoffs","Four families of one-shot federated learning, decoded","One-shot FL survey: generative models beat distillation","A hands-on taxonomy for one-shot federated learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000336,"raw_usage":{"total_tokens":1798,"prompt_tokens":820,"completion_tokens":978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":436,"completion_tokens_details":{"reasoning_tokens":896}},"tokens_in":436,"tokens_out":978,"duration_ms":8921,"temperature":1.0,"reasoning_tokens":896,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T22:35:12.647926+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the surveyed methods on a single fixed benchmark with the same datasets, model architectures, compute budgets, and privacy settings; if generative models do not systematically beat knowledge distillation, or adaptive ensembles do not beat static ones, the paper's central comparative findings fail.","supporting_citations":[{"cited_title":"Federated learning with matched averaging","cited_arxiv_id":null,"evidence_quote":"Extends prototype matching to CNNs and LSTMs, forming the FedMA backbone within parameter learning."},{"cited_title":"Dense: Data-free one-shot federated learning","cited_arxiv_id":null,"evidence_quote":"Dense is the data-free generative-plus-ensemble-plus-distillation baseline that later hybrid methods measure against."},{"cited_title":"Enhancing one- shot federated learning through data and ensemble co- boosting","cited_arxiv_id":null,"evidence_quote":"Co-Boosting provides evidence for the finding that iterative server-side updating after one upload improves accuracy."},{"cited_title":"Revisiting ensembling in one-shot federated learn- ing","cited_arxiv_id":null,"evidence_quote":"FENS provides evidence for the claim that adaptive ensemble weighting beats static ensembling."}],"review_version":1}