{"id":"e1aff782-8a0a-417b-a90f-f2b660e22944","arxiv_id":"2501.10463","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"GLow implements and simulates fully decentralized gossip learning in the Flower framework, achieving competitive accuracy with centralized and federated baselines on MNIST and CIFAR10.","lead":"GLow is a new simulation strategy that lets researchers run gossip-style decentralized machine learning using the Flower framework, with no central server. It reaches 98% accuracy on MNIST and 75% on CIFAR10 in small-network experiments, and the authors say it matches centralized and federated learning closely.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GLow is not actually an asynchronous gossip simulator: Algorithm 1 is a globally synchronized round-robin head loop, and Section VI admits future parallel work, so the deployment-assessment claim exceeds what the tool currently simulates.","rationale":"The reader's conditional verdict is appropriate, but the sharpest reason is more specific than missing packet-loss or Byzantine modeling. The paper's own definitions and limitation statements make Algorithm 1 a synchronous round-robin serialization of a single-head aggregation step, not an asynchronous gossip process. Section II.C characterizes gossip learning as asynchronous exchange; Section VI admits the current simulation sequentially selects one head and only projects a parallel version. Since the abstract's value proposition is assessing devices before physical deployment, the simulator should at least emulate the concurrency and stochastic activation of real gossip systems. It currently cannot, so the reported results do not transfer to the target setting. Separately, Table II's own 16+4 CIFAR10 numbers (0.683 versus 0.778/0.789) undercut the 'similar in all designed experiments' sentence, and the post-hoc selection of the double-ring topology for the headline numbers weakens the comparative evidence further. I would not reject the paper: the code is public, the special-agent methodology is a useful primitive, and the serialized experiments are internally coherent. I would keep the paper conditional, requiring either an asynchronous execution mode with a comparative evaluation or a rescoping of all deployment-related claims, and requiring all topologies with error bars rather than a single selected topology. The proposed asynchronous comparison would settle whether the current results are an artifact of the serialization or a genuine property of the GLow strategy.","tokens_in":13064,"tokens_out":9926,"duration_ms":108201,"concrete_test":"Implement an asynchronous execution mode in GLow: replace the global 'for iteration' loop of Algorithm 1 with per-agent random activation (e.g., Poisson clocks) and pairwise pull/aggregate operations, with no global round-robin or global iteration counter; keep the same topologies, epochs, data splits, and aggregation rule. Run the 8+2 and 16+4 MNIST and CIFAR10 scenarios for topologies 2, 4, and 7 with at least 5 seeds, and compare mean and per-agent accuracy/convergence curves to the current synchronous mode. If accuracy differs by more than the seed-level error bars (e.g., >0.02) or convergence rounds shift systematically, then Algorithm 1's results are specific to the synchronous serialization and the deployment-assessment claim must be rescoped.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is more specific than missing packet-loss or Byzantine fault modeling: the simulation semantics do not match the target system. Section II.C defines Gossip Learning as agents exchanging parameters asynchronously without a server, but Algorithm 1 runs a global loop with head k = iteration mod K, aggregates neighbors' current weights, and repeats. Section VI explicitly states the current GLow 'is designed to sequentially select an agent as head' and that a 'parallel version' is only projected. This is a synchronous, single-head, round-robin schedule, not an asynchronous peer-to-peer gossip process. Consequently, the central claim that GLow 'will allow researchers to train and assess scalability and convergence ... before making a physical deployment' is unsupported for the class of asynchronous gossip systems the paper motivates. The reported MNIST/CIFAR10 curves characterize only this serialized variant; deployments with concurrent agent activity, message delays, or stragglers may converge differently. The empirical claim cannot rescue the paper either: Table II reports 0.683 for GLow on 16+4 CIFAR10 versus 0.778 (FL) and 0.789 (CNL), which contradicts 'performs similarly ... in all designed experiments,' and only the post-hoc double-ring topology is reported. Both issues need resolution before the headline capability claim is credible.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents GLow, a simulation-oriented Gossip Learning (GL) strategy built on the Flower framework. GLow designates one agent as a head in each iteration via a round-robin schedule (Algorithm 1); the head trains locally and aggregates the current weights of its neighbors. A topology generator and special-agent configurations (disconnected, empty, or both) are introduced. The system is evaluated on MNIST and CIFAR10 with 8+2 and 16+4 agents and compared against centralized learning (CNL) and FedAvg (FL). The authors report that GLow achieves competitive accuracy in the 8+2 scenario and in the MNIST 16+4 scenario, but acknowledge lower accuracy for CIFAR10 with 16+4 agents.","tokens_in":13366,"tokens_out":6609,"duration_ms":66754,"significance":"If the claims are validated, a flexible, open-source, Flower-based GL simulator would be a useful community tool for studying decentralized learning before deployment. The authors provide a public repository and visualize per-agent behavior, which is a strength. However, the paper's central contribution is weakened by a mismatch between the simulated scheduling and the asynchronous gossip model introduced in the paper, by an overstatement in the abstract relative to Table II, and by a communication-round comparison that is not controlled across systems. The special-agent methodology is a useful experimental idea, though it is used more as a sanity check than as a substantive comparison.","major_comments":[{"comment":"The paper defines Gossip Learning as a process where agents exchange parameters asynchronously without a server. However, Algorithm 1 implements a globally synchronized loop that selects k = iteration mod K as the single head, aggregates neighbor weights, and repeats. This is a synchronous, single-head, round-robin schedule, not an asynchronous peer-to-peer gossip process. Section VI confirms that the current version 'is designed to sequentially select an agent as head' and that a parallel version is only projected. Consequently, the abstract's claim that GLow 'will allow researchers to train and assess scalability and convergence of devices ... before making a physical deployment' is not supported for the asynchronous gossip systems the paper motivates. The authors should either implement asynchronous event-driven simulation or explicitly limit the claims to the synchronous variant.","section":"Section II.C and Algorithm 1"},{"comment":"The abstract states that GLow 'performs similarly in terms of accuracy and convergence to its analogous Centralized and Federated approaches in all designed experiments.' Table II contradicts this: for CIFAR10 with 16+4 agents, GLow achieves 0.683, while FL achieves 0.778 and CNL achieves 0.789, a gap of roughly 0.1. The conclusions soften this to 'slightly lower accuracies,' but 0.1 is substantial. The abstract and conclusions should be revised to accurately reflect the 16+4 CIFAR10 result, or the experiment should be re-run with more communication rounds or different hyperparameters to close the gap.","section":"Table II and Abstract"},{"comment":"The communication-round comparison is not apples-to-apples. In FL, a communication round is a global round in which all clients train and the server aggregates. In GLow, Section IV.C defines a communication round as the number of times each agent is designated as head, so the total number of head aggregations is agents × rounds (e.g., 8 × 24 = 192 for MNIST 8+2). Thus GLow performs many more aggregation events than FL for the same reported 'Communication Rounds' value, while the local-epoch count per agent is similar. The convergence comparison should be controlled for the number of aggregation rounds or message exchanges, or the discrepancy should be explicitly stated and justified.","section":"Section IV.C and Table II"},{"comment":"The empirical evaluation only compares GLow against centralized learning and vanilla FedAvg. Since the paper positions GLow as a GL simulator, it should also compare against existing decentralized baselines such as the gossip learning of Hegedus et al. [2], GossipFL [14], BrainTorrent [11], or Fedstellar [12]. Without such baselines, the claim that GLow is a useful GL strategy is not strongly established. At minimum, the authors should discuss how the synchronous round-robin schedule differs from these baselines and why the comparison to FedAvg alone is sufficient for the paper's claims.","section":"Section V"}],"minor_comments":[{"comment":"The text contains a typo: 'Aditionally' should be 'Additionally'.","section":"Section V.C"},{"comment":"The subfigure label '(d) Topo 7 - F. connected' should be spelled out as 'Fully connected' for clarity.","section":"Figure 1"},{"comment":"The model is attributed to 'Lecun et Al.' in the text, but the standard spelling is 'LeCun et al.' The paper should use consistent naming.","section":"Section IV.A"},{"comment":"The text states that GLow maintains modularity by modifying only upper-level classes, but the footnote says a 'slight modification in the internal aggregate class of server' was made. This contradiction should be clarified.","section":"Section III"},{"comment":"The table reports 'Average Accuracy' for GLow only for topology 4 (double ring), while figures show results for many topologies. The selection criterion for reporting topology 4 should be stated explicitly in the table caption or text.","section":"Table II"},{"comment":"The expected random-guessing behavior of ED agents is used as a validation of the system, but this expectation is trivially true for agents with no data and no connections. The paper should not present this as a strong evidence of correctness.","section":"Section V.B"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as a systems/tooling contribution rather than an algorithmic advance. The main technical issues (synchronous vs. asynchronous schedule, communication-round mismatch, and the 16+4 CIFAR10 result) can be fixed by reframing the claims and adding appropriate baselines, but the current abstract and conclusions overstate the findings. I would support publication after major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work with Flower or want to prototype decentralized FL. The gap is real: Flower is server-centric and no one has published a simulation-oriented, fully decentralized strategy on top of it, so GLow fills that niche. The topology generator and the 'special agents' methodology (empty, disconnected, etc.) are genuinely useful ideas, and the code is public. The MNIST numbers are fine; the 8+2 CIFAR10 result is also competitive.\n\nBut the headline claims need scoping. First, Algorithm 1 is a globally synchronized round-robin head loop, not asynchronous peer-to-peer gossip. The paper motivates GL with asynchrony, and Section VI admits a parallel version is only planned, so the abstract's promise that researchers can 'assess scalability and convergence ... before making a physical deployment' is stronger than what the simulator currently delivers. Second, the abstract says GLow 'performs similarly' in all designed experiments, but Table II shows 0.683 on 16+4 CIFAR10 vs 0.778 for FL and 0.789 for CNL. That is a real gap, and the discussion in Section V-C partly acknowledges it, but the abstract should be toned down. Third, the empirical section is thin: no error bars, and Table II reports only the double-ring topology, chosen after the fact as the first topology where E and R converge similarly. That selection is defensible, but it weakens the comparative claim.\n\nNone of this is fatal. The paper is honest about the sequential head limitation in the conclusions, the related work is well handled, and the central message—that a decentralized Flower strategy can be built at the strategy layer—holds up. What needs fixing is claim-scoping and a bit more statistical support. I'd send it to peer review: a serious referee could get it into decent shape, and the community would benefit from the tool.","headline":"Useful Flower-based decentralized learning simulator, but the 'performs similarly' claim overreaches and the synchronous round-robin design means it is not yet an asynchronous gossip simulator.","tokens_in":13860,"tokens_out":1808,"would_cite":false,"duration_ms":18309,"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":"GLow builds a Flower-based gossip learning simulator that matches federated accuracy in tests.","keywords":["gossip learning","federated learning","decentralized learning","Flower framework","network topology","simulation","MNIST","CIFAR10"],"falsifier":"Run GLow's double-ring 8+2 agent CIFAR10 configuration on physical devices or in an emulator with random message drops, delayed arrivals, or a malicious agent sending corrupted weights; if the connected agents' average accuracy drops substantially below the disconnected self-learning baseline, the simulator's convergence predictions do not transfer to real conditions.","tokens_in":1334,"feed_emoji":"🌸","tokens_out":3532,"duration_ms":73091,"temperature":0.7,"pith_summary":"This paper introduces GLow, a gossip learning strategy built on the Flower framework, which simulates fully decentralized training without a central server. The authors claim GLow lets researchers test convergence, scalability, and topology choices for IoT-style agent networks before physical deployment. Across MNIST and CIFAR10, with 8 and 16 connected agents plus special control agents, GLow reaches accuracies over 0.98 and 0.75, and performs similarly to centralized and federated learning in most experiments. The point is that fully decentralized gossip learning can be studied and validated in simulation before dealing with the complexity and faults of real devices.","feed_headline":"Flower-based gossip learning matches federated accuracy in simulation","feed_subtitle":"GLow lets researchers test custom topologies and convergence before any physical deployment.","key_machinery":"The central mechanism is a round-robin head-aggregation loop: each iteration selects one agent as head, that agent trains locally for a fixed number of epochs, pulls current weights from its neighbors, and replaces its model with a weighted average of those weights. This replaces the server-driven parameter spread of a centralized federated strategy with a neighbor-request step, reducing the number of messages. A topology generator creates graphs from disconnected to fully connected, and special agents (empty, disconnected, empty-and-disconnected) act as controls to verify that information propagates through the network.","core_discovery":"The central claim is that a round-robin head aggregation scheme, implemented as a custom Flower strategy, reproduces the accuracy and convergence behavior of centralized and federated approaches in fully decentralized settings. In each communication round the designated head agent trains locally for E epochs, pulls current weights from its neighbors, and sets its model to a weighted average of those weights. The paper shows that with sufficient connectivity, roughly a double-ring topology, agents without any local data converge to the same accuracy as data-holding neighbors, while disconnected agents learn only from their own data and perform worse. The authors conclude that GLow is competitive with vanilla FedAvg and centralized learning, despite using only a subset of the data and including data-less agents in the network.","pith_inferences":["Because GLow's simulation assumes synchronous rounds, reliable message delivery, and no Byzantine or packet-loss behavior, the strong accuracy results may not transfer to real deployments; extending the simulator to inject message drops, delays, or malicious agents would test this directly.","The saturation of accuracy beyond a double-ring topology suggests a general principle: sparse connectivity may suffice for gossip-style decentralized learning, which could guide communication budget choices in real systems.","The round-robin scheduling scheme can be seen as a deterministic traversal of the agent graph; comparing it with random head selection or multiple simultaneous heads would isolate the effect of scheduling on convergence and fairness.","The special-agent methodology (empty and disconnected agents as controls) could be reused in other decentralized learning simulators as a lightweight way to diagnose whether learned information actually propagates or merely reflects local data."],"forward_implications":["Researchers can use GLow to estimate the convergence and accuracy of custom gossip topologies before physical deployment, under the simulator's reliable-communication assumptions.","Adding connectivity beyond a double ring yields little accuracy gain but can speed up early convergence, suggesting a sparse-connectivity design point for decentralized systems.","Including data-less agents exposes whether information truly propagates through the network; with sufficient connectivity these agents converge to neighbor-level accuracy.","In the 8+2 agent CIFAR10 scenario GLow reaches 0.754 accuracy versus 0.791 for federated averaging, a small gap given the removed server bottleneck; scaling to 16+4 agents lowers CIFAR10 accuracy to 0.683, attributed to data splitting and disconnected agents.","GLow outperforms self-learning in every connected topology tested, showing that decentralized parameter averaging provides a measurable benefit over isolated training."],"supporting_citations":[{"why":"Defines FedAvg, the federated baseline against which GLow's accuracy and convergence are compared.","marker":"[1]"},{"why":"Supplies the Flower simulation framework that GLow modifies at the strategy level.","marker":"[3]"},{"why":"Introduces gossip learning with linear models, the decentralized training paradigm GLow simulates.","marker":"[8]"},{"why":"Provides the MNIST handwritten-digit dataset used for the image classification experiments.","marker":"[16]"},{"why":"Provides the CIFAR10 dataset used for the more complex image classification experiments.","marker":"[17]"},{"why":"Lecun et al.'s LeNet architecture used as the model for both datasets.","marker":"[18]"}],"fun_headline_variants":["Flower-based gossip learning matches federated accuracy in simulation","GLow simulates gossip learning to match federated accuracy without a central server","Gossip learning simulation via Flower matches federated accuracy","GLow: Flower-based gossip learning matches federated accuracy","Simulating gossip learning with Flower matches federated accuracy"],"cache_read_input_tokens":16000,"weakest_assumption_plain":"The simulation assumes communication is synchronous, reliable, and free of Byzantine or packet-loss behavior, so accuracy measured in GLow may not hold in real decentralized deployments where these failures occur.","fun_headline_variants_meta":{"raw":{"variants":["Flower-based gossip learning matches federated accuracy in simulation","GLow simulates gossip learning to match federated accuracy without a central server","Gossip learning simulation via Flower matches federated accuracy","GLow: Flower-based gossip learning matches federated accuracy","Simulating gossip learning with Flower matches federated accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3512,"prompt_tokens":868,"completion_tokens":2644,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":2560}},"tokens_in":484,"tokens_out":2644,"duration_ms":18091,"temperature":1.0,"reasoning_tokens":2560,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:17:45.170283+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GLow's double-ring 8+2 agent CIFAR10 configuration on physical devices or in an emulator with random message drops, delayed arrivals, or a malicious agent sending corrupted weights; if the connected agents' average accuracy drops substantially below the disconnected self-learning baseline, the simulator's convergence predictions do not transfer to real conditions.","supporting_citations":[{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Defines FedAvg, the federated baseline against which GLow's accuracy and convergence are compared."},{"cited_title":"Gossip learning with linear models on fully distributed data,","cited_arxiv_id":null,"evidence_quote":"Introduces gossip learning with linear models, the decentralized training paradigm GLow simulates."}],"review_version":1}