{"id":"96b6883a-8cce-497d-a7e8-32fa4256b607","arxiv_id":"2502.09624","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":7,"one_line_summary":"A graph Resfusion model that starts from a feasible route and denoises it toward a shorter route achieves lower block propagation time and higher trust scores than greedy, genetic, GAT, and Difusco baselines in simulated blockchain-enabled vehicle networks.","lead":"This paper combines a trust-scoring method with a graph-based diffusion model to choose the path a new blockchain block should travel through a fleet of mobile AI devices. The authors report faster and more trustworthy block delivery than four routing baselines in simulations on real taxi-trajectory datasets from Chengdu and Porto.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training labels optimize Euclidean tour length, but Eq. (22) defines the objective as Shannon transmission time; since the two objectives can disagree, the model is not trained for the stated optimization problem.","rationale":"I read the paper as a system-paper application of a generative diffusion model to block propagation in consortium blockchain-enabled MEANETs. The strongest elements are the clear assembly of a trust-cloud scoring mechanism, a warm-started graph Resfusion model, and an extensive simulation comparison on two real trajectory datasets. The central claim is that the proposed model generates the optimal block propagation trajectory with respect to the stated objective. The most load-bearing point is that the model is trained with labels from a Euclidean TSP solver, while Eq. (22) defines the cost as a nonlinear function of distance, namely Shannon transmission time. Since a strictly increasing nonlinear re-weighting of edge costs can change which Hamiltonian cycle is optimal, the training signal does not match the evaluation objective unless the two optima are shown to coincide, which the paper does not do. This is an internal correctness concern rather than a disagreement with external consensus, and it is directly testable with the paper's own equations and datasets. The reader identified the PBFT-to-TSP sequential-tour premise as the weakest assumption; I agree that premise is also serious, but the Euclidean-label versus Shannon-objective mismatch is more decisive because it affects the claimed optimality even if one accepts the TSP formulation. The reader's conditional verdict remains appropriate: the concern is reparable by regenerating labels with Shannon costs and rerunning the comparisons, but as submitted the central optimality result is not established.","tokens_in":18906,"tokens_out":7551,"duration_ms":77107,"concrete_test":"Take 100 random instances with V = 58, 78, 98, 118 from the same trajectory datasets; for each, compute (a) the Concorde Euclidean-optimal tour used as the training label, and (b) a Concorde-optimal tour under the Shannon edge costs of Eq. (22) using the same coordinates and Table II parameters. If the Shannon-optimal tour has lower Eq. (22) cost than the Euclidean-optimal tour by any nontrivial margin (e.g., >1%) on a substantial fraction of instances, the training labels do not match the stated objective and the central optimality claim fails. A second check is to retrain the graph Resfusion model on Shannon-cost labels and compare the resulting Fig. 6 curves; a material change in reported propagation times would confirm the mismatch.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V-A defines the objective as minimizing total block propagation time, with edge cost Λ_{i-1,i} = S_block / (W log2(1 + aρΓ_{i-1,i}^{-ε}/(N0W))) (Eq. (22)). Section VI-A2 states that training labels are generated by the Concorde exact solver following [28]; in [28] the labels are Euclidean TSP tours, and Section VI-B and Fig. 11 evaluate trajectory lengths in a unit square. Thus the graph Resfusion model is trained to minimize sum of distances, not sum of Shannon times. Because Λ(d) is strictly increasing but not affine in distance, the Hamiltonian cycle minimizing total Euclidean length need not minimize Eq. (22). The paper gives no argument that the two optima coincide for the simulated Chengdu and Porto geometry. Consequently, the generated 'optimal block propagation trajectory' is optimal for a different objective, and the efficiency comparisons in Fig. 6 do not establish the claimed optimality. The Section III-A premise that PBFT block propagation is a sequential TSP tour is also unverified, but the training-label mismatch is internal and directly testable with the paper's own equations.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for trustworthy and efficient block propagation in consortium blockchain-enabled Mobile Embodied AI Networks. A cloud-model-based trust score filters out low-trust miners, and a graph Resfusion model (a diffusion model with residual guidance, using a GatedGNN) is trained to generate a block propagation trajectory. The optimization objective is stated as minimizing total Shannon-capacity transmission time subject to a trust-score threshold. The paper reports simulations on Chengdu and Porto vehicle-trajectory datasets, comparing against greedy, genetic algorithm, attention-model-based, and Difusco baselines. The central claims are that the proposed method generates near-optimal propagation trajectories, reduces failed samples compared with Difusco, and adapts well to dynamic topologies.","tokens_in":19188,"tokens_out":4026,"duration_ms":42529,"significance":"If the central claims are substantiated, the paper would offer a practical combination of trust filtering and diffusion-based combinatorial optimization for a relevant mobile networking problem. The adaptation of Resfusion to a graph setting, the use of real urban trajectory data, and the explicit specification of channel parameters are useful strengths. The training-and-sampling pipeline is clearly described, and the generalization test across problem sizes is a valuable check. However, the claimed optimality with respect to the stated Shannon-time objective is not established by the current training-label design, and the underlying sequential-TSP model of PBFT propagation is not justified. The paper also does not provide code, statistical error bars, or external validation of the trust model. These issues are load-bearing for the main efficiency and trustworthiness claims.","major_comments":[{"comment":"The training labels do not match the optimization objective. Eq. (22) defines the cost to minimize as the sum of Shannon transmission times Lambda_{i-1,i} = S_block / (W log2(1 + a rho Gamma_{i-1,i}^{-epsilon}/(N0 W))), which is strictly increasing but nonlinear in distance. Section VI-A2 says labels are generated by the Concorde exact solver following [28], and Section VI-B/Fig. 11 evaluate trajectory lengths in a unit square. Thus the model is trained to minimize Euclidean tour length, not the Shannon-time sum of Eq. (22). Since edge costs in Eq. (22) are heterogeneous and non-affine functions of distance, the minimizing Hamiltonian cycle for Euclidean length need not minimize the stated objective. The paper gives no argument that the two optima coincide for the Chengdu and Porto geometries, so the reported 'optimal block propagation trajectory' and the efficiency comparisons in Fig. 6 do not establish optimality for the problem the paper formulates.","section":"Section V-A and Section VI-A2"},{"comment":"The sequential-TSP premise for PBFT block propagation is unverified and load-bearing. Section III-A states that after filtering untrusted miners, 'the remaining propagation process can be modeled as the traveling salesman problem (TSP) in terms of miner node selection alone [22],' with the leader sequentially forwarding the block to each miner and collecting votes. Practical Byzantine Fault Tolerance normally requires broadcast or all-to-all message exchanges among replicas in its prepare and commit phases; a single sequential tour through each miner does not obviously implement PBFT. The cited reference [22] addresses consortium blockchain energy trading and does not provide a PBFT message-level model. If this assumption is false, the optimized trajectory is for a different propagation mechanism and the framework's applicability to PBFT-based consortium blockchains is not established. The authors should specify the PBFT message exchanges this sequential tour implements, or explicitly adopt a different consensus protocol.","section":"Section III-A"},{"comment":"The empirical outperformance claim is not accompanied by any uncertainty quantification. Fig. 6 reports point estimates only for the total block propagation time, yet the compared methods include stochastic algorithms (GA, attention model, Difusco, and the proposed diffusion-based method). Without error bars, multiple seeds, or significance tests, the stated performance ranking (proposed > Difusco > GAT > GA > Greedy) is not statistically supported. The failure-probability comparison in Section VI-E does use 32 parallel samplings, but the propagation-time comparison does not report any such replication; this asymmetry weakens the main efficiency conclusion.","section":"Section VI-B and Fig. 6"},{"comment":"The reduced failure probability claimed in Fig. 10 is partly forced by construction, so it does not by itself demonstrate higher inference accuracy. As shown in Eq. (31), the reverse Resfusion process starts from G_T' = sqrt(alpha_T') G_hat_0 + sqrt(1 - alpha_T') epsilon, i.e., a noisy version of the already-feasible degraded trajectory G_hat_0. Compared with Difusco, which starts from pure noise, the Resfusion sampler is initialized in a feasible region and will trivially produce fewer infeasible trajectories. The paper should separate this initialization effect from genuine solution-quality gains; for example, report constraint satisfaction conditioned on equal-distance-from-feasible-set starting points, and report the achieved objective value of the decoded trajectories, not only their feasibility.","section":"Eq. (31) and Section VI-E"}],"minor_comments":[{"comment":"There are repeated typographical errors, including 'Resfuion model' in Section VI-B and VI-E, and 'graph Resfuion model' in Section VI-E; these should be corrected.","section":"Throughout"},{"comment":"The baseline called 'Graph attention network' is cited as [38], but [38] is Kool et al., 'Attention, learn to solve routing problems!', which is an attention model for routing, not a graph attention network in the usual GAT sense. The citation and the algorithm description should be aligned.","section":"Section VI-B"},{"comment":"The generalization table reports percentages in a matrix but does not define the exact formula used to compute them or the baseline convention (the diagonal is labeled as the benchmark). Please state the computation explicitly in the text or caption so the reader can interpret the 'maximum percentage does not exceed 2%' claim.","section":"Section VI-E / Fig. 11"},{"comment":"The trustworthiness capability indicators are described ambiguously: 'average miner block propagation time' is listed as a positive ability indicator, while lower propagation time is usually preferable. If the intended indicator is throughput or a normalized delay metric, this should be stated to avoid confusion.","section":"Section IV-C"},{"comment":"The notation T' in Algorithm 1 is computed by arg min over i of |sqrt(alpha_i) - 1/2|, but alpha_i is not defined in the algorithm block; it should reference Eq. (27) or define the cumulative product explicitly. Also, Algorithm 1 computes T' but then samples t uniformly from {1,...,T'}; the role of T' in training should be explained.","section":"Section V-D and Algorithm 1"},{"comment":"The manuscript has no limitations paragraph. Given the load-bearing modeling assumption in Section III-A and the training-label issue in Section VI-A2, a brief discussion of the scope and threats to validity would improve the paper.","section":"Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent adaptation of Resfusion/Difusco to a block-propagation scenario, but the mismatch between the Euclidean-TSP training labels and the Shannon-time objective is a fundamental correctness issue for the stated optimality claim. The PBFT sequential-tour assumption also needs explicit justification. Both are fixable within the scope of a revision, which is why I recommend major revision rather than rejection. I would also ask the editor to require that the authors either release code or provide enough implementation detail for the label-generation and evaluation pipeline to be independently reproduced; the current manuscript gives no code availability statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent incremental systems paper. It assembles three existing pieces, trust-cloud scoring, GatedGNN, and Resfusion, into a block-propagation routing heuristic for consortium-blockchain MEANETs. The writing is clear, the authors honestly credit Resfusion and Difusco, and the simulation comparison covers four baselines on two real taxi trajectory datasets. That part is fine, and the paper deserves a referee slot because the combination is new and the problem is real.\n\nThe soft spots are not fatal to the paper as a heuristic, but they do undercut the title claim of optimality. Section V-A states the objective as Shannon-capacity transmission time (Eq. 22); training labels are generated with Concorde following [28], which solves Euclidean TSP. Since the time cost is strictly increasing but nonlinear in distance, a minimum-length tour need not minimize the time sum. The paper gives no argument that the optima coincide for the Chengdu and Porto geometries. The stress-test note is correct: the model is trained for a different objective from the one being reported. This is an internal mismatch, directly testable from the paper's own equations.\n\nThe second load-bearing assumption is the PBFT-to-TSP model in Section III-A. PBFT normally uses all-to-all or broadcast message exchange; sequential forwarding as a TSP tour is a strong simplification. The paper cites [22] for this, but it is not verified against the actual PBFT message pattern. If that assumption fails, the optimized trajectory does not correspond to how blocks actually propagate. This is a systems-modeling concern, and it may be fixable by framing the work as a routing heuristic for a sequential gossip protocol rather than as PBFT itself.\n\nMinor issues: no error bars on the main efficiency plots, no code release, and the 19.60% trust-score improvement mainly shows that filtering changes the score, not that it measurably improves security. The 'first to consider trustworthiness' claim is also weakened by [8] from the same group, which already considers reputation in block propagation.\n\nWho this is for: readers working on blockchain propagation for vehicular networks, or researchers applying diffusion models to routing. It is an incremental contribution, not a breakthrough. With a re-framed objective, a check on the PBFT assumption, and a reproducibility appendix, it could become a solid systems paper. As is, it deserves serious peer review, but the referee should push hard on the training-objective mismatch.","headline":"A competent incremental systems paper whose central optimality claim is undercut by training on Euclidean TSP while reporting a Shannon-capacity time objective; it deserves a serious referee but needs a re-framed objective and a check on the PBFT-to-TSP assumption.","tokens_in":19727,"tokens_out":1717,"would_cite":false,"duration_ms":16278,"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 a graph Resfusion model can compute the fastest block-propagation route through trusted miners in consortium blockchain-enabled MEANETs, improving both efficiency and trustworthiness.","keywords":["block propagation","consortium blockchain","mobile embodied AI networks","trust cloud model","graph diffusion model","Resfusion","gated graph neural network","traveling salesman problem"],"falsifier":"Simulate or instrument a real PBFT-based consortium blockchain in which the leader broadcasts the block to all replicas rather than touring them sequentially; if the graph Resfusion route does not reduce the measured end-to-end confirmation time under that message pattern, the claimed optimal trajectory does not correspond to actual propagation. A second check is to run the model on standard TSP instances with known optimal values and compare the returned tour lengths; if the gap is no better than a simple 2-opt baseline, the efficiency claim is not supported.","tokens_in":18699,"feed_emoji":"⛓","tokens_out":8866,"duration_ms":70527,"temperature":0.7,"pith_summary":"This paper tries to establish that block propagation in consortium blockchains serving mobile embodied AI networks (MEANETs) can be made simultaneously faster and more trustworthy by planning the propagation route with a graph-based diffusion model. The proposed framework first computes a trust score for each miner from reputation, trustworthiness, and risk clouds, then feeds an easily available but unoptimized propagation route into a graph Resfusion model that denoises it into a near-optimal trajectory under a trust constraint. In simulations on real vehicle trajectories, the model is reported to beat a greedy mechanism, a genetic algorithm, a graph attention network, and a diffusion-based combinatorial solver in propagation time, while raising the total trust score of participating miners by 19.60% at 99 vehicles. If the result holds, fleet operators get a practical way to keep blockchain latency low while excluding malicious or unreliable nodes.","feed_headline":"Trusted-miner routes cut block propagation delay in MEANETs","feed_subtitle":"Graph Resfusion filters untrusted miners and finds the fastest trusted block route in simulations.","key_machinery":"The load-bearing object is the graph Resfusion model: a diffusion model whose forward process corrupts the optimal trajectory with Gaussian noise plus a residual $R = \\hat{G}_0 - G_0$, and whose reverse process, parameterized by a GatedGNN, estimates the residual noise $res^\\epsilon$ to recover $G_0$ from a degraded trajectory. The GatedGNN is an anisotropic graph network that updates edge embeddings alongside node embeddings, letting it score candidate edges of the propagation tour. A second load-bearing component is the cloud-model trust score, computed as $TS_{r_j} = (S_{rep} + S_{tw} - e^{RiskC(S_{rep})} - e^{RiskC(S_{tw})})/(2e)$, which filters out miners below the threshold $\\lambda$ before routing.","core_discovery":"On the paper's own terms, the central discovery is that the block-propagation routing problem can be solved by a diffusion model that learns to remove both noise and a known residual. The optimization objective is to minimize the total propagation time $\\sum_{i=2}^K \\Lambda_{i-1,i}$ subject to each selected miner's trust score $TS_i > \\lambda$, where propagation delay follows a Shannon-formula rate. The graph Resfusion model treats the optimal propagation trajectory $G_0$ as the target, defines a degraded trajectory $\\hat{G}_0$ that is easy to construct, forms the residual $R = \\hat{G}_0 - G_0$, and learns to predict the residual noise $res^\\epsilon$ with a GatedGNN. The reverse process starts from a computable trajectory at step $T'$ instead of pure noise, which the authors argue reduces sampling space and inference time. Simulations show shorter propagation time than the baselines and a 33.3% lower failure rate than Difusco at 99 vehicles.","pith_inferences":["We infer the same residue-constrained diffusion recipe could be applied to other sequential routing problems on graphs, such as UAV inspection planning or delivery routing, wherever a cheap feasible tour and a learned residual denoiser are available.","We infer the trust-cloud scoring is protocol-agnostic and could be reused in other PBFT-style consortium chains, but the TSP premise would need to be revalidated for any consensus protocol that broadcasts in parallel.","A natural testable extension is to replace the sequential tour with a parallel broadcast tree or DAG and let the GatedGNN predict edges of that structure; the residual-resfusion mechanism should transfer directly."],"forward_implications":["If the framework is correct, consortium blockchains in MEANETs can propagate blocks over shorter, high-trust routes, which lowers the chance of forks and signature-collection failures.","Filtering by the cloud-model trust score raises the average trust of participating miners, reducing exposure to selfish mining and Sybil attacks; the paper reports a 19.60% trust-score gain at 99 vehicles.","The residual-guided reverse process reaches a valid trajectory in 368 diffusion steps instead of the 1000 steps of a standard diffusion model, which matters for dynamically changing topologies.","Generalization tests in the paper show trajectory-length deviation below 2% when the model is evaluated on fleet sizes different from training, suggesting retraining may not be needed after every topology change."],"supporting_citations":[{"why":"Supplies the Resfusion residual-noise diffusion process and the reduced-step reverse starting point.","marker":"[24]"},{"why":"Provides the graph-diffusion combinatorial solver that the paper extends and the baseline it must beat.","marker":"[28]"},{"why":"Provides the premise that block propagation after trust filtering is a traveling salesman problem.","marker":"[22]"},{"why":"Supplies the reputation, trustworthiness, and risk formulas used in the trust-score calculation.","marker":"[23]"},{"why":"Introduces the trust cloud model that gives the randomness-and-fuzziness trust evaluation.","marker":"[14]"},{"why":"Provides the prior GAT-based block propagation scheme and the channel parameter settings used in simulation.","marker":"[8]"},{"why":"Serves as the genetic algorithm baseline for propagation-time comparison.","marker":"[37]"},{"why":"Serves as the graph attention network baseline for propagation-time comparison.","marker":"[38]"},{"why":"Supplies the GatedGNN edge and node update equations used to estimate residual noise.","marker":"[36]"}],"fun_headline_variants":["Graph Resfusion cuts block propagation delay in MEANETs","Resfusion model routes blocks via trusted miners only","Diffusion-based block routing trims delay and failure","Blockchain MEANETs: Resfusion improves propagation trust","Trusted block paths via graph Resfusion in MEANETs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire optimization rests on the premise that after untrusted miners are filtered out, PBFT-style block propagation in a consortium blockchain is a sequential tour: the leader forwards the block to one trusted miner at a time, collects each vote, and returns, so the routing problem is a traveling salesman problem.","fun_headline_variants_meta":{"raw":{"variants":["Graph Resfusion cuts block propagation delay in MEANETs","Resfusion model routes blocks via trusted miners only","Diffusion-based block routing trims delay and failure","Blockchain MEANETs: Resfusion improves propagation trust","Trusted block paths via graph Resfusion in MEANETs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000243,"raw_usage":{"total_tokens":1569,"prompt_tokens":1026,"completion_tokens":543,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":642,"completion_tokens_details":{"reasoning_tokens":461}},"tokens_in":642,"tokens_out":543,"duration_ms":5259,"temperature":1.0,"reasoning_tokens":461,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T14:19:25.572185+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate or instrument a real PBFT-based consortium blockchain in which the leader broadcasts the block to all replicas rather than touring them sequentially; if the graph Resfusion route does not reduce the measured end-to-end confirmation time under that message pattern, the claimed optimal trajectory does not correspond to actual propagation. A second check is to run the model on standard TSP instances with known optimal values and compare the returned tour lengths; if the gap is no better than a simple 2-opt baseline, the efficiency claim is not supported.","supporting_citations":[{"cited_title":"Resfusion: Denoising Diffusion Probabilistic Models for Image Restoration Based on Prior Residual Noise","cited_arxiv_id":"2311.14900","evidence_quote":"Supplies the Resfusion residual-noise diffusion process and the reduced-step reverse starting point."},{"cited_title":"Difusco: Graph-based diffusion solvers for combinatorial optimization,","cited_arxiv_id":null,"evidence_quote":"Provides the graph-diffusion combinatorial solver that the paper extends and the baseline it must beat."},{"cited_title":"Consortium blockchain for secure energy trading in industrial Internet of Things,","cited_arxiv_id":null,"evidence_quote":"Provides the premise that block propagation after trust filtering is a traveling salesman problem."},{"cited_title":"Trust mechanism-based multi-tier computing system for service-oriented edge-cloud networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the reputation, trustworthiness, and risk formulas used in the trust-score calculation."},{"cited_title":"An intelligent trust cloud management method for secure clustering in 5G enabled Internet of Medical Things,","cited_arxiv_id":null,"evidence_quote":"Introduces the trust cloud model that gives the randomness-and-fuzziness trust evaluation."},{"cited_title":"Graph attention network-based block propagation with optimal AoB and reputation in Web 3.0,","cited_arxiv_id":null,"evidence_quote":"Provides the prior GAT-based block propagation scheme and the channel parameter settings used in simulation."},{"cited_title":"Genetic algorithm,","cited_arxiv_id":null,"evidence_quote":"Serves as the genetic algorithm baseline for propagation-time comparison."},{"cited_title":"Attention, learn to solve routing problems!","cited_arxiv_id":null,"evidence_quote":"Serves as the graph attention network baseline for propagation-time comparison."}],"review_version":1}