{"id":"35cae533-0847-450b-9acd-5bd430be65be","arxiv_id":"2504.16732","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A blockchain-free peer-to-peer swarm learning framework shows modest accuracy on a three-class histopathology task, but missing details and weak baselines undermine the claim of parity with centralized models.","lead":"This paper proposes a simplified peer-to-peer swarm learning framework that drops blockchain and shares model updates directly between nodes, tested on cancer histopathology images. The authors report the approach performs close to centralized training while keeping patient data local, though the experimental evidence is incomplete.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's central claim of 'comparable performance' is contradicted by its own reported numbers, and the peer-aggregation rule is left unspecified.","rationale":"The reader's weakest_assumption—that the validation-threshold peer aggregation converges to a useful consensus on non-IID data—is indeed the most load-bearing unsupported step. My independent reading confirms this and adds that the paper's own reported numbers refute the 'comparable performance' claim, since Node 3's best swarm result is only 80% of the centralized AUC. The manuscript also lacks a dataset name, a concrete aggregation formula, a convergence analysis, and any ablation of the acceptance threshold. These are not mere stylistic gaps; they make the empirical results non-reproducible and the central claim unsubstantiated. The reader's verdict of REJECT with moderate confidence is appropriate. I agree with the reader's identification of the weakest assumption, and my check of the numerical evidence further strengthens the rejection.","tokens_in":5402,"tokens_out":1094,"duration_ms":11387,"concrete_test":"Request that the authors provide the exact aggregation update rule (equation), the dataset identity, and the precise data split used for Node 0. Then re-run the reported P2P-SL experiment on that dataset with the stated rule and check whether Node 3's swarm AUC reaches at least 0.7156 (matching centralized) or whether the paper's own numbers hold; if the aggregation formula or dataset cannot be disclosed, the central empirical claim is unverifiable.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim (Abstract: 'achieving comparable performance to centralized models') is undercut by the reported results. Node 3's swarm model reaches AUC 0.6892, only 80% of the centralized 0.7156 baseline (Section 4.2), a gap the paper itself frames as partial recovery, not comparability. Node 0 improves from 0.6192 to 0.6397, still far below baseline, and with overlapping error bars (±0.0057 vs ±0.0036) the gain is not shown to be statistically meaningful. More importantly, the load-bearing mechanism—the peer aggregation rule described as 'weighted averaging of LoRA-adapter weights accepted on a validation AUC threshold of 80%' (Sections 3.2 and 4.5)—is never given a concrete formula. No aggregation equation, no convergence analysis, no ablation of the threshold, and no dataset name are provided. The paper's empirical claims cannot be reproduced or independently checked without these details. The single simulated split with five seeds does not support robustness claims on non-IID data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a blockchain-free peer-to-peer swarm learning (P2P-SL) framework for decentralized training on cancer histopathology images. The method replaces blockchain-based consensus with direct peer exchanges of LoRA-adapter weights, using a validation-based acceptance threshold. Experiments compare swarm-trained models with a centralized baseline and fully local models across four nodes with imbalanced data. The central claim is that P2P-SL achieves performance comparable to centralized training while preserving privacy and improving robustness against overfitting and data imbalance.","tokens_in":5693,"tokens_out":4029,"duration_ms":35306,"significance":"If the central claim were supported, the work would be a useful step toward practical decentralized learning for privacy-sensitive medical imaging. The motivation is reasonable: removing the blockchain layer from swarm learning could lower deployment barriers. The evaluation design, with multiple seeds, a held-out test set, and comparisons to centralized and local baselines, is a sensible start. However, the reported numbers do not support the 'comparable performance' claim, and the core aggregation mechanism is not specified precisely enough to reproduce. The paper also leaves the dataset unnamed, which prevents independent verification. As presented, the contribution is not yet established at a level suitable for publication.","major_comments":[{"comment":"The central claim of 'comparable performance to centralized models' is contradicted by the reported results. The centralized baseline reaches AUC 0.7156, while the swarm model on Node 0 attains only 0.6397±0.0036 and the swarm model on Node 3 attains 0.6892±0.0063. Even the best swarm node remains below the baseline, and the paper's own description of Node 3 as 'recovered over 80% of centralized performance' is not the same as comparable performance. Because the primary contribution is this performance claim, the manuscript's central result is unsupported by its own evidence.","section":"Abstract and Section 4.2"},{"comment":"The model architecture description is internally inconsistent. The text states that the encoder modules reduce dimensionality to 1024, but the next sentence says a fully connected layer transforms features 'from 1152 to 512 dimensions.' No explanation reconciles 1024 and 1152. Additionally, the final layer maps to three dimensions with a sigmoid activation, whereas a three-class histopathology classification task typically requires a softmax over mutually exclusive classes or an explicit multi-label formulation. These inconsistencies make the model impossible to reproduce as written.","section":"Section 3.3"},{"comment":"The peer aggregation rule, which is the load-bearing mechanism of the proposed framework, is never specified. The paper says only that the framework uses 'weighted averaging of LoRA-adapter weights accepted on a validation AUC threshold of 80%,' but gives no formula for the weights, no description of how the 80% threshold is applied (e.g., per-node local validation or global), and no convergence analysis or ablation of the threshold. Without a concrete aggregation rule, the experimental results cannot be reproduced, and the claim that the method is 'robust' lacks a theoretical or empirical basis.","section":"Sections 3.2 and 4.5"},{"comment":"The evaluation dataset is not identified. The paper mentions '10,000 annotated histopathology images' with Macenko stain normalization and augmentation, but it does not name the dataset, the tissue type, the number of classes, or the class distribution. Since the model is initialized from TorchXRayVision, a model pre-trained on chest X-rays, the reader cannot judge whether the transfer is plausible. The absence of a dataset name and split details makes the entire empirical study irreproducible.","section":"Section 4.1"},{"comment":"Several robustness claims are presented without the supporting data needed to evaluate them. The manuscript reports a 50% reduction in generalization gap, a 35% decrease in cross-validation variance, a 15% lower Davies-Bouldin Index, a 4.5% improvement in minority-class recall, and an 8% reduction in false positives and negatives, but none of these figures is accompanied by error bars, raw values, or a statistical test. The t-SNE analysis is mentioned without a corresponding figure. These unsupported quantitative claims should be either substantiated or removed.","section":"Sections 4.3 and 4.4"}],"minor_comments":[{"comment":"The phrase 'recovered over 80% of centralized performance' is imprecise: 0.6892/0.7156 is approximately 0.96, so 'over 80%' is technically true but does not quantify how close the swarm model actually is to the baseline.","section":"Section 4.2"},{"comment":"The listed keywords ('Single-cell Sequencing Integration · Multi-Omics · Dimensionality Reduction · Normalization') are unrelated to the paper's content and appear to be a copy-paste error.","section":"Keywords"},{"comment":"The text contains a LaTeX-to-text artifact: 'training AUC ¿ 0.95' should read '> 0.95'.","section":"Section 4.4"},{"comment":"Several references are incomplete or malformed. For example, reference [12] lacks a publisher, year, and page range, and reference [9] has an awkward line break in the URL.","section":"References"},{"comment":"The t-SNE analysis is described in prose but no t-SNE figure or quantitative embedding evaluation is shown, making the claim about 'tighter intra-class clustering' difficult to verify.","section":"Section 4.3"},{"comment":"The term 'federated-average unbalanced' is used to describe the experimental configuration, but the paper proposes swarm learning, not federated learning; consistent terminology would avoid confusion.","section":"Section 4.1"}],"recommendation":"reject","confidential_remarks":"The paper's core claim is contradicted by its own reported AUCs, and the aggregation mechanism is left unspecified. The unnamed dataset and missing architecture details make the study irreproducible. These are load-bearing issues that would require new experiments and a full rewriting of the method description to address, rather than local revisions. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper describes P2P-SL, a simplified swarm learning framework that removes the blockchain layer and has peers exchange LoRA adapter weights every three epochs, gated by a validation-AUC threshold of 80%. Tested on histopathology classification with four nodes. The engineering is plausible and the literature review is fair: it cites Saldanha et al. and Warnat-Herresthal et al. and positions itself as a simplification, not a new paradigm. The experiments use five seeds, a held-out test set, and include t-SNE and generalization-gap analysis, which is decent practice.\n\nThe main problem is not the arithmetic of the headline claim. Node 3's swarm AUC is 0.6892 against centralized 0.7156, which is about 96% of baseline, not 'only 80%' as one quick reading suggests. Node 0 reaches 0.6397, about 89%, which is a real gap. So 'comparable performance' is defensible for Node 3 and a stretch for Node 0, but it is not flatly contradicted by the numbers. The real soft spots are incompleteness and inconsistency. The aggregation rule is described as 'weighted averaging' with an 80% validation threshold, but no formula, no definition of the weights, and no acceptance criterion are given. The dataset is never named—'10,000 annotated histopathology images' is all we get. The architecture section says the encoder reduces to 1024 dimensions, then a fully connected layer maps 1152 to 512; that is a concrete inconsistency. And sigmoid activation on a three-class output is unusual unless this is multi-label, which is not explained. No code is provided, so none of this is checkable.\n\nThe robustness claims rest on one simulated split and an ablation-free threshold choice. The peer-exchange interval and threshold are free parameters without sensitivity analysis. These are not fatal to the idea, but they are fatal to the current draft's ability to support the conclusions.\n\nI'd send this to peer review rather than desk-reject it. The system is real, the simplification of swarm learning is a legitimate question, and the defects are omissions and unclear reporting, not a fundamentally broken approach. A serious referee could ask for the missing formula, dataset name, and code, and the paper might be salvageable. For my own reading group, I'd bring it only as a case study in how to (and not to) report distributed ML experiments.","headline":"A plausible P2P swarm-learning system with honest citations and decent experimental practice, but the aggregation rule and dataset are underspecified and the 'comparable performance' claim is a stretch for Node 0.","tokens_in":6135,"tokens_out":4872,"would_cite":false,"duration_ms":42588,"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 claims that removing the blockchain layer from swarm learning—replacing it with direct peer exchange of small model updates—still yields accuracy close to a centrally trained diagnostic model while keeping patient images on-site.","keywords":["swarm learning","peer-to-peer","histopathology","cancer diagnostics","decentralized training","model aggregation","privacy preservation","imbalanced data"],"falsifier":"Run the same four-node setup on a public histopathology dataset with a held-out multi-institution test set: if peer-exchanged models do not beat locally trained models on the held-out set, or if a 30%-data node fails to reach roughly 80% of the centralized baseline's AUC, the central claim fails. A second check: scaling to 10 or more nodes should not collapse performance; if accuracy degrades sharply with swarm size, the aggregation rule is not robust as claimed.","tokens_in":5272,"feed_emoji":"🩺","tokens_out":5674,"duration_ms":50446,"temperature":0.7,"pith_summary":"The paper tests a stripped-down version of swarm learning for cancer histopathology: nodes keep their images local, train on their own data, and periodically exchange small model updates with peers instead of relying on a blockchain to coordinate. The central claim is that this simplified peer-to-peer design is enough to approach the accuracy of a model trained on all data pooled centrally, while also reducing overfitting. On the reported split, the data-poor Node 0 moves from an AUC of 0.6192 alone to 0.6397 with peer exchange, and Node 3 reaches 0.6892, over 80% of the centralized baseline's 0.7156. If the claim holds, it means small institutions with little data and no blockchain infrastructure could still participate in collaborative diagnostic-model training.","feed_headline":"Swarm learning without blockchain matches a centralized cancer-diagnosis model","feed_subtitle":"A peer-to-peer version lifts a 10%-data node's AUC from 0.619 to 0.640 while keeping images on-site.","key_machinery":"The load-bearing mechanism is threshold-gated weighted averaging of low-rank adapter weights exchanged over a peer-to-peer network. Each node trains locally on its own images, shares only the small adapter updates every three epochs, and applies an update only if it passes a validation AUC threshold of 80%, so aggregation is adaptive rather than dictated by a central ledger. The vision pipeline is a pre-trained encoder originally built for chest X-rays plus a dense convolutional decoder, which the paper says transfers well to histopathology. This mechanism carries the argument because all claimed gains—privacy, scarce-data recovery, and overfitting reduction—follow from local training plus selective peer-weight exchange.","core_discovery":"The paper's discovery claim is that a decentralized learning system can be simplified to a direct peer exchange of small low-rank adapter weights, with each node accepting a peer's contribution only when a validation-based threshold (80% AUC) is met, and that this reproduces most of the benefit of centralized training. In the federated-average unbalanced experiment, local training on 10% of the data yields AUC 0.6192; adding peer exchange every three epochs raises that to 0.6397, while a node holding 30% reaches 0.6892 versus a centralized baseline of 0.7156. The paper also reports that swarm-trained models have a smaller train-validation generalization gap than local models, which it attributes to aggregation acting as an implicit regularizer.","pith_inferences":["The paper's evidence is one simulated split on an unspecified dataset; a natural extension is to test whether the 80% validation threshold and three-epoch exchange interval transfer across datasets, node counts, and non-IID severity.","If threshold-gated averaging works because it filters harmful updates, it connects to the broader model-merging literature, where more sophisticated trajectory- or uncertainty-based weighting exists; the simple threshold is a minimal version of that idea.","A concrete testable extension is to vary the acceptance threshold and measure AUC: if the threshold has a large effect, it is the true control parameter; if not, the benefit may come from exchange frequency or dataset size."],"forward_implications":["If the framework works as described, blockchain is not a necessary component of swarm learning; removing it lowers the barrier for non-technical or under-resourced sites.","A node holding only 10% of the data can improve its diagnostic AUC by roughly 0.02, and a 30%-data node can recover over 80% of the centralized baseline, so participation becomes useful even for small datasets.","The reduced generalization gap reported under peer aggregation suggests the exchange step can act as a regularizer, which would matter for other small-sample medical imaging tasks.","Because only adapter weights are exchanged, communication cost is low, making the scheme compatible with modest bandwidth in real hospital networks."],"supporting_citations":[{"why":"Defines the federated averaging rule that the paper's weighted local aggregation adapts.","marker":"[9]"},{"why":"Shows swarm learning applied to cancer histopathology, the target domain this work simplifies.","marker":"[13]"},{"why":"Establishes the privacy-preserving swarm-learning setting and the blockchain-based architecture the paper removes.","marker":"[15]"},{"why":"Motivates validation- or uncertainty-based model merging, which the 80% acceptance threshold instantiates in simplified form.","marker":"[6]"},{"why":"Surveys aggregation techniques in federated learning, framing the baseline the framework must match.","marker":"[12]"}],"fun_headline_variants":["Swarm learning without blockchain rivals centralized cancer diagnosis","Peer-to-peer swarm learning lifts small-node cancer AUC","Blockchain-free swarm learning achieves comparable cancer diagnosis","Simplified P2P swarm learning rivals centralized cancer AI","Decentralized peer swarm learning narrows gap to central model"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's success rests on the assumption that averaging low-rank adapter weights that pass an 80% validation threshold will keep improving each node's model on unseen data; the paper offers no convergence analysis and only one simulated split with one dataset as evidence.","fun_headline_variants_meta":{"raw":{"variants":["Swarm learning without blockchain rivals centralized cancer diagnosis","Peer-to-peer swarm learning lifts small-node cancer AUC","Blockchain-free swarm learning achieves comparable cancer diagnosis","Simplified P2P swarm learning rivals centralized cancer AI","Decentralized peer swarm learning narrows gap to central model"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000707,"raw_usage":{"total_tokens":3156,"prompt_tokens":883,"completion_tokens":2273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":2196}},"tokens_in":499,"tokens_out":2273,"duration_ms":18233,"temperature":1.0,"reasoning_tokens":2196,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:56:32.488954+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same four-node setup on a public histopathology dataset with a held-out multi-institution test set: if peer-exchanged models do not beat locally trained models on the held-out set, or if a 30%-data node fails to reach roughly 80% of the centralized baseline's AUC, the central claim fails. A second check: scaling to 10 or more nodes should not collapse performance; if accuracy degrades sharply with swarm size, the aggregation rule is not robust as claimed.","supporting_citations":[{"cited_title":"Nature medicine 28(6), 1232–1239 (2022)","cited_arxiv_id":null,"evidence_quote":"Shows swarm learning applied to cancer histopathology, the target domain this work simplifies."},{"cited_title":"Nature 594(7862), 265–270 (2021)","cited_arxiv_id":null,"evidence_quote":"Establishes the privacy-preserving swarm-learning setting and the blockchain-based architecture the paper removes."},{"cited_title":"In: 2022 2nd International Conference on Advance Computing and Innovative Technologies in Engineering (ICACITE)","cited_arxiv_id":null,"evidence_quote":"Surveys aggregation techniques in federated learning, framing the baseline the framework must match."}],"review_version":1}