{"id":"1872775f-040f-4406-ada1-afb45293f735","arxiv_id":"2505.11760","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Topology-aware aggregation using degree or betweenness centrality improves out-of-distribution knowledge propagation in decentralized learning across 36 simulated topologies and five datasets.","lead":"Researchers study how out-of-distribution (OOD) knowledge spreads across devices in decentralized learning, and find it spreads much worse than regular data. They propose weighting neighbors by network centrality during model aggregation, which improves OOD accuracy by 123% on average in simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'OOD knowledge' claim relies on a single backdoor trigger; degree-weighted aggregation may spread that artifact without generalizing to realistic distribution shifts.","rationale":"The reader's CONDITIONAL verdict and moderate confidence match my assessment. The paper is internally consistent: under its stated definition of OOD, the experiments plausibly show that backdoor knowledge spreads poorly under topology-unaware aggregation and better under topology-aware aggregation. The evidence base is broad in terms of datasets, topologies, and seeds, but every result is generated through one artificial OOD proxy. The 123% improvement figure is not derived explicitly in the text, and even if it is an accurate summary of Fig. 4, it quantifies backdoor-trigger propagation, not general out-of-distribution knowledge. Missing heterogeneous baselines such as RelaySum and absent error bars are real reporting gaps, but they affect precision and comparison fairness rather than the core conceptual claim. The load-bearing weakness is construct validity: the paper argues about OOD knowledge but measures backdoor memorization, and the two need not share the same propagation dynamics. A conditional acceptance requiring at least one realistic OOD test with a matched sample-size control is therefore the appropriate decision.","tokens_in":24019,"tokens_out":4112,"duration_ms":43918,"concrete_test":"Run the Section 5.1 protocol (BA topologies with p in {1,2,3}, 33 nodes, CIFAR10, R=40, E=5, same seeds) with a non-backdoor OOD shift: train all nodes on CIFAR10 classes 0 through 7, place classes 8 and 9 only on the highest-degree node, and evaluate every node on classes 8 and 9. Compare OOD AUC for Degree, Betweenness, and Unweighted. To control for sample size, repeat with the OOD node holding the same number of OOD samples as a typical IID node holds IID samples. If Degree/Betweenness no longer outperform Unweighted by a comparable margin, the headline claim is backdoor-specific rather than a general property of OOD knowledge propagation.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central empirical claim — that OOD knowledge is harder to propagate and that Degree/Betweenness aggregation accelerates it — is measured entirely through one backdoor operationalization. Appendix B.2.2 defines OOD as relabeling 10% of one node's images with a red square (or token-forcing in TinyMem), and every OOD accuracy, AUC, and the 123% abstract figure is accuracy on backdoored test inputs. A backdoored model can acquire the trigger by learning a small, highly salient visual pattern with a fixed target label; this is closer to memorizing a conspicuous artifact than to acquiring a general OOD concept such as a new class, covariate shift, or semantic shift. Increasing the aggregation weight of high-degree nodes may spread that artifact quickly while having a different effect on knowledge that requires more capacity or more diverse samples to encode. In addition, the OOD data are smaller in volume than the IID data on a node (10% of one node's images vs. the node's full IID set), so part of the observed IID/OOD AUC gap in Fig. 2 may be a sample-size effect rather than an OOD effect. If the mechanism is 'rare, distinctive pattern from a central source spreads via hubs', the claim is correct but narrow; if the intended claim is general topology-aware knowledge dissemination, the current evidence does not establish it. The paper's own limitation section discusses malicious-data spread but does not address proxy validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies decentralized learning over static communication topologies, where each node trains locally and aggregates with neighbors. It claims that out-of-distribution (OOD) knowledge is systematically harder to propagate than IID knowledge, that propagation depends on OOD data location and topology, and that topology-aware aggregation weights based on degree or betweenness centrality (softmax-scaled with temperature tau) accelerate OOD knowledge propagation, improving OOD test accuracy by 123% on average relative to topology-unaware baselines. The evidence base is broad: five datasets, three topology families, 36 topologies, four OOD locations, and three seeds, with code released on GitHub. The OOD notion is operationalized exclusively as a backdoor trigger (a red square on one node's images, or token-forcing in a language task; Appendix B.2.2).","tokens_in":24270,"tokens_out":5479,"duration_ms":55087,"significance":"If the central claim holds, the paper delivers a simple and practically appealing result: replacing uniform or data-size aggregation weights with graph-centrality-based weights can substantially improve dissemination of non-IID knowledge in decentralized systems with no degradation of IID accuracy. The study is unusually broad in topology and dataset coverage, and the authors release code and report energy/carbon accounting, which supports reproducibility. However, significance is currently tempered by three gaps: the OOD phenomenon is measured through a single backdoor artifact, the IID-vs-OOD comparison is confounded with single-source placement, and the strongest non-IID baselines are absent. Each of these gaps narrows the scope of the claim that 'OOD knowledge' generally propagates poorly and that the proposed methods beat existing decentralized learning strategies.","major_comments":[{"comment":"The operationalization of OOD knowledge is a single backdoor trigger: a red n x n square relabeled to a fixed class (Def. B.1) or token-forcing (Def. B.2). All OOD accuracy, AUC, and the abstract's 123% figure measure the model's accuracy on backdoored test inputs only. A backdoor is a highly salient, low-capacity artifact, and its propagation may follow different dynamics than the propagation of a new class, a covariate shift, or a semantic shift. The paper's title and abstract make a general claim about OOD knowledge, but the evidence supports only 'backdoor-trigger knowledge.' To make the central claim load-bearing, the authors should either (a) add at least one alternative OOD construction (e.g., a new class present on a single node, or a style/rotation shift), or (b) explicitly narrow the claims and title to backdoor knowledge propagation. This is the paper's key correctness-risk point.","section":"Section 3, Appendix B.2.2"},{"comment":"The comparison of IID vs OOD propagation is confounded by data placement. IID data are distributed across all nodes (Section B.2.1, alpha_l = alpha_s = 1000), whereas the OOD data are placed on a single node only (Section B.2.2, Q = 10% of one node's data). Thus the OOD knowledge must travel from one source while the IID knowledge is already present everywhere; the observed gap in Fig 2 may reflect single-source propagation rather than an intrinsic difficulty of OOD knowledge. The paper should include a control condition in which an equally sized 'IID' dataset is also placed on a single node (or, conversely, OOD data are spread across all nodes). Without this control, the first contribution ('OOD knowledge is more difficult to propagate') is not established.","section":"Section 3, Fig 2, Appendix B.2.1"},{"comment":"The baseline set omits strong non-IID methods, most notably RelaySum, which is cited in Related Work as [74] and is specifically designed for decentralized learning on heterogeneous data. RelaySum is topology-aware in the sense that it weights neighbors based on the rooted-tree distance, so it is the natural competitor to test the claim that 'topology-aware aggregation strategies outperform ... existing decentralized learning strategies' (abstract, contribution 3). The four baselines used (Unweighted, Weighted, Random, FL) are all topology-naive for the non-IID setting. Adding RelaySum (and, if feasible, other non-IID baselines such as consensus-control methods) is needed before the paper can claim superiority over existing strategies. Alternatively, the claims should be restricted to comparisons against these particular baselines.","section":"Section 5.1, Appendix B.3, Related Work [74]"},{"comment":"The headline improvement of '123% on average across models in a topology' is not defined anywhere. There is no equation or description of how the percentage is computed: the numerator and denominator (e.g., relative increase in OOD AUC of Degree vs Unweighted?), the averaging set (across datasets? topologies? seeds? nodes?), and whether it is a mean of per-node percentages or a ratio of means. Moreover, all plotted results (Figs 4, 5, 6) show only three-seed means with no error bars, standard deviations, or significance tests. Given the centrality of this number and the small number of seeds, the authors must provide the exact calculation and some measure of variability. This is a load-bearing quantitative claim and cannot be checked in its current form.","section":"Abstract, Section 5.1, Fig 4"}],"minor_comments":[{"comment":"The first sentence says 'This experiment is identical to that of Section 4'; it should refer to Section 5.1, since Section 4 is the proposed method, not an experiment.","section":"Section 5.2"},{"comment":"The y-axis 'OOD vs IID AUC % Diff.' is not defined; the text should state whether this is (AUC_IID - AUC_OOD)/AUC_IID or another formula, and confirm it is averaged over nodes and then over topologies.","section":"Fig 2"},{"comment":"The Random baseline is defined with a softmax temperature tau, but the experiments (e.g., Section 5.1) do not report the value of tau used for Random; please state it explicitly and, if it is set to 0.1, say so.","section":"Appendix B.3, item 3"},{"comment":"There is a typo: 'prorogation' should be 'propagation.'","section":"Section 5.1, 'Result & Discussion'"},{"comment":"For the TinyMem language dataset, the evaluation metric is described as testing whether the next tokens match the backdoored suffix, but it is not explained how this is converted into the 'OOD AUC' values plotted in Figs 4-5; please clarify the scoring procedure and how accuracy is accumulated over rounds.","section":"Appendix B.2.2, Table 1, Figs 4-5"},{"comment":"The heatmap axes for the row/column strategies are not labeled in the figure panels; labels such as 'OOD node rank' and aggregation strategy names would make the numeric tables interpretable without cross-checking the caption and main text.","section":"Appendix C.2, Figs 15-18"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical study with broad coverage, but its central claim outruns the experiments in three ways: the single-backdoor OOD proxy, the missing RelaySum-style baseline, and the undefined 123% headline. These are fixable within the manuscript's scope if the authors add control experiments and tighten the claims. The paper may also be a better fit for a systems or applied-ML venue than a theory venue. No ethical concerns; the energy/carbon reporting is a plus."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth a serious referee. It asks whether weighting aggregation coefficients by graph centrality accelerates propagation of OOD knowledge in decentralized learning, and answers with a broad simulation campaign: five datasets, BA/SB/WS topologies, three seeds, and a simple method—softmax over degree or betweenness centrality. The finding that OOD knowledge is harder to propagate than IID knowledge, and that centrality weighting helps, is credible within the paper's own operationalization.\n\nThe weighting idea itself is not new; the paper cites centrality-aware gossiping and topology-impact studies. What is new is the systematic evaluation in deep decentralized learning with OOD propagation as the target, and the observation that both data location and topology matter in a way a drop-in weighting can partially fix. The paper also reports energy and carbon costs and links to code, which is more than many submissions do.\n\nThe load-bearing weakness is the OOD proxy. All OOD data is a single backdoor type—a red square or token-forcing—and every OOD accuracy, AUC, and the 123% abstract figure is accuracy on backdoored test inputs. That measures how quickly a conspicuous artifact spreads, not how well a general OOD concept such as a new class or covariate shift is acquired. The stress-test note about sample size is also fair: 10% of one node's data is far smaller than that node's IID set, so part of the IID/OOD AUC gap in Figure 2 may reflect volume rather than OOD-ness. The limitations section acknowledges that the method propagates malicious data but does not address whether backdoors represent realistic distribution shifts.\n\nSecond, the evaluation omits strong non-IID baselines like RelaySum, and results are reported as averages without error bars or significance tests. The headline 123% improvement appears in the abstract, but I could not find the calculation in the text. That is a reporting gap, not evidence of anything worse, but it matters for a claim this prominent.\n\nThe paper's own limitation statement is honest: the methods only diverge from Unweighted when node-level characteristics vary, and they propagate malicious data too. That suggests the authors know the boundaries of their approach.\n\nOverall, the central claim is plausible but narrower than the title suggests. With a second OOD operationalization, RelaySum added, variance reported, and the 123% figure derived, the stronger claim would be supportable. As written, it is a solid conditional accept. Send it to peer review; a serious referee can push on the right points and the paper is honest enough to benefit.","headline":"A credible but narrow empirical study: centrality-weighted aggregation speeds propagation of a backdoor-style OOD signal, and the paper deserves peer review with serious pressure on the OOD proxy and missing baselines.","tokens_in":24788,"tokens_out":2043,"would_cite":true,"duration_ms":21642,"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":"In decentralized learning, out-of-distribution knowledge propagates poorly along arbitrary network topologies; weighting aggregation by degree or betweenness centrality fixes much of the gap, raising OOD accuracy 123% on average.","keywords":["decentralized learning","knowledge propagation","out-of-distribution generalization","network topology","aggregation weights","graph centrality","backdoor triggers"],"falsifier":"Take the same protocol but replace the backdoor trigger with a natural OOD shift—for example, hold out one CIFAR-10 class entirely and place it on a single node as its local data—then measure OOD test accuracy across all nodes under Degree/Betweenness versus Unweighted aggregation. If the topology-aware advantage of 123% does not appear (or shrinks to noise), the claim that centrality-weighted aggregation accelerates general OOD knowledge propagation is falsified.","tokens_in":23794,"feed_emoji":"🌐","tokens_out":7349,"duration_ms":64606,"temperature":0.7,"pith_summary":"Decentralized learning trains a separate model on each device in a network, and models learn about distant data only by repeated averaging with their immediate neighbors. The paper argues that out-of-distribution (OOD) knowledge—knowledge that is rare or foreign to most devices—is systematically harder to push through such a topology than ordinary IID knowledge, and that how well it spreads depends on where in the graph the OOD data live and on the graph's own structure. It proposes a small change to the aggregation rule: weight each neighbor's model by that neighbor's degree or betweenness centrality, scaled by a softmax temperature, rather than weighting all neighbors equally. In experiments with five datasets and 36 topologies, these topology-aware rules improve OOD test accuracy by 123% on average over topology-unaware baselines, without hurting IID accuracy. A sympathetic reader would take away that a purely local, graph-informed choice of averaging weights can close much of the OOD propagation gap in decentralized learning.","feed_headline":"123% better OOD spread: weight neighbors by network centrality","feed_subtitle":"A simple change to aggregation weights closes the OOD propagation gap without hurting IID accuracy.","key_machinery":"The mechanism is the aggregation step of decentralized learning: $m^{t+1}_i \\leftarrow \\sum_{j \\in N_i} C_{i,j} m^{t+\\frac12}_j$, where $C_{i,j}$ is the weight device $i$ places on neighbor $j$'s model. Normally $C_{i,j}$ is uniform, proportional to data count, or random. The paper's innovation is to set $C_{i,j} = e^{R_j/\\tau} / \\sum_{k \\in N_i} e^{R_k/\\tau}$, with $R_j$ the degree or betweenness centrality of neighbor $j$ and temperature $\\tau=0.1$ in the experiments. This single coefficient choice is what carries the argument: it makes hubs (or bridges) absorb and re-emit more knowledge each round, accelerating the hop-by-hop spread of the OOD signal through the topology.","core_discovery":"The central discovery is that OOD knowledge does not propagate through decentralized topologies the way IID knowledge does: when a single node holds data that is out of distribution for the rest of the network, models on nodes far from that origin learn the OOD signal far worse after 40 aggregation rounds, even though IID accuracy is roughly uniform across the topology. The location of the OOD data matters—placing it on high-degree nodes spreads it much further than placing it on low-degree nodes—and so does the topology: higher average degree helps, higher modularity (tightly clustered communities) hurts. The paper's proposed remedy is to replace uniform or data-count-based aggregation weights with weights derived from graph centrality: each node aggregates its neighbors' models in proportion to $e^{R_j/\\tau}$ where $R_j$ is neighbor $j$'s degree or betweenness centrality and $\\tau$ is a temperature. With these topology-aware weights, OOD test accuracy rises by 123% on average across models in a topology compared to topology-unaware baselines, and IID accuracy is not degraded.","pith_inferences":["If the result generalizes beyond the backdoor proxy—say a genuinely new class or a natural distribution shift introduced at one device—then centrality-weighted aggregation could serve as a communication-only fix for OOD generalization in edge networks; the paper does not test this directly.","The reported metric is accuracy AUC over 40 rounds, which averages final level and arrival time; a natural test is to measure how many rounds are needed to reach a threshold OOD accuracy, since centrality weighting plausibly accelerates the early phase of propagation even when final AUC is similar.","The softmax temperature $\\tau$ is fixed at 0.1 in all experiments; varying $\\tau$ interpolates between uniform averaging ($\\tau\\to\\infty$) and winner-take-all ($\\tau\\to0$), so a sweep could reveal whether the 123% gain is robust or concentrated in this one operating point."],"forward_implications":["The aggregation change costs no extra communication at training time: each device needs only precomputed centrality values for its neighbors, so adopting the strategy is a drop-in replacement for uniform averaging.","OOD knowledge seeded on high-degree nodes spreads much further than on low-degree nodes under every strategy, so placing rare or safety-critical data on well-connected devices is itself a propagation lever.","Higher modularity (tight communities) slows OOD spread; topology-aware weights shrink but do not eliminate that penalty, so operators of community-structured networks should expect more rounds before distant devices learn rare knowledge.","Because the same weighting also speeds propagation of malicious or erroneous model updates, safeguards for detecting unwanted knowledge should accompany any deployment."],"supporting_citations":[{"why":"Supplies the Barabási-Albert model used to generate the scale-free topologies in the main experiments.","marker":"[5]"},{"why":"Defines betweenness centrality, one of the two topology-aware aggregation metrics.","marker":"[16]"},{"why":"Provides the single-target image backdoor design used to create OOD training and test data.","marker":"[18]"},{"why":"Provides the stochastic block model used to generate modular topologies for the modularity experiments.","marker":"[23]"},{"why":"Supplies the decentralized SGD framework with local updates and changing topology on which the experiments are based.","marker":"[36]"},{"why":"Established that a node's learning performance under heterogeneous data depends on graph centrality; this paper extends that observation into aggregation design.","marker":"[61]"},{"why":"Provides the TinyMem dataset and the token-forcing backdoor recipe used for the language experiments.","marker":"[68]"},{"why":"Provides the Watts-Strogatz small-world model used to test node-count effects.","marker":"[78]"}],"fun_headline_variants":["Centrality-based aggregation boosts OOD spread by 123%","Weight neighbors by centrality for 123% OOD gain","Topology-aware weights improve OOD knowledge propagation","123% OOD boost from centrality-weighted aggregation","Graph centrality aggregation spreads OOD knowledge better"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The experiments assume that a single device holding 10% backdoored data (a red-square trigger on images, or a forced token sequence in language) is a faithful stand-in for out-of-distribution knowledge, and that accuracy on backdoored test data measures how well that knowledge spread; if real OOD knowledge behaves differently than such an easily detectable trigger, the conclusions may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Centrality-based aggregation boosts OOD spread by 123%","Weight neighbors by centrality for 123% OOD gain","Topology-aware weights improve OOD knowledge propagation","123% OOD boost from centrality-weighted aggregation","Graph centrality aggregation spreads OOD knowledge better"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000197,"raw_usage":{"total_tokens":1353,"prompt_tokens":920,"completion_tokens":433,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":358}},"tokens_in":536,"tokens_out":433,"duration_ms":4494,"temperature":1.0,"reasoning_tokens":358,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:48:21.571092+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same protocol but replace the backdoor trigger with a natural OOD shift—for example, hold out one CIFAR-10 class entirely and place it on a single node as its local data—then measure OOD test accuracy across all nodes under Degree/Betweenness versus Unweighted aggregation. If the topology-aware advantage of 123% does not appear (or shrinks to noise), the claim that centrality-weighted aggregation accelerates general OOD knowledge propagation is falsified.","supporting_citations":[{"cited_title":"Impact of network topology on the performance of Decentralized Federated Learning","cited_arxiv_id":"2402.18606","evidence_quote":"Established that a node's learning performance under heterogeneous data depends on graph centrality; this paper extends that observation into aggregation design."},{"cited_title":"Mitigating memorization in language models","cited_arxiv_id":null,"evidence_quote":"Provides the TinyMem dataset and the token-forcing backdoor recipe used for the language experiments."}],"review_version":1}