{"id":"4ea906f0-0cd2-4471-a900-2fe6994098d0","arxiv_id":"2601.15944","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An Ising-model message-passing algorithm partitions large networks into candidate synchronized clusters, recovers topological-equivalence groups, and shows abrupt desynchronization that smooths into plateaus under noise.","lead":"This paper uses a message-passing (belief propagation) algorithm on an effective Ising model to partition large networks into clusters of nodes that would move in step. The goal is a scalable, dynamics-agnostic probe for synchronization clusters without choosing a specific microscopic model.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Noisy variant internally inconsistent: continuous random fields (Sec. VI.C) plus machine-precision cluster threshold (Eq. 18) should give Q=N and p=0 for all J, contradicting Figs. 4–6; reported plateaus require an undocumented tolerance.","rationale":"I focused on the internal inconsistency between continuous random fields and the machine-precision clustering threshold because it is a concrete, checkable contradiction in the stated algorithm, and it undermines the noisy version that is used for the WordNet application and for the paper's noise-smoothing conclusions. The reader's formal weakest_assumption was the surrogate-dynamics validity, so my agreement is partial, although the reader did mention this threshold issue in the rationale. If confirmed, the noisy results are not reproducible as written, and the paper needs to state and justify the actual clustering tolerance or re-run the simulations. The noiseless version may still be salvageable, so the verdict remains conditional rather than accept or reject; no change to the reader's CONDITIONAL verdict is needed.","tokens_in":17022,"tokens_out":7747,"duration_ms":96269,"concrete_test":"Inspect the archived Fortran code (Zenodo DOI 10.5281/zenodo.18235744) to find the numerical value of ε used in the clustering step of the noisy version; then re-run the noisy simulation of Fig. 4/5 with ε set to 10^-15 and the same random seed. If Q=N and p=0 for all J, the noisy results are artifacts of a larger tolerance. If the code already uses a larger tolerance, document that value and test whether the plateaus persist under changes of that tolerance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Sec. VI.C, the random fields h_i are drawn independently from a continuous uniform distribution. For generic h_i, the local magnetizations computed from Eq. (10) are distinct for every node. The clustering step Eq. (18) groups nodes only when |x_i-x_j|<ε, and Sec. VI states that ε is chosen equal to machine precision (10^-15 or 10^-18). Under these stated parameters, no two magnetizations should coincide, so Q=N and p=0 for every J. Yet Figs. 4–6 show non-trivial Q and p with well-defined plateaus. Either the code uses a much larger ε in Eq. (18), in which case the partition and all plateau features depend on an arbitrary tolerance that is never disclosed, or the reported noisy results cannot be reproduced from the stated algorithm. This directly affects the WordNet application, which uses only the noisy version, and the conclusion that noise smooths the transition. The noiseless method may be salvageable, but the noisy variant as written does not satisfy the paper's own target definition in Sec. III. The paper nowhere explains this contradiction, and the code is the only way to resolve it.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a message-passing algorithm (MPA) on a binary Ising model with uniform couplings as an 'unbiased surrogate dynamics' for detecting clusters of synchronized nodes in large networks. Three variants are presented: noiseless with positive initial conditions, noiseless with random initial conditions, and a noisy version with random local fields. The method is validated on a synthetic graph with known topologically equivalent (TE) groups and applied to the US power-grid and WordNet networks. The paper reports abrupt desynchronization events in the noiseless case and smoother, plateau-like behavior with noise, and emphasizes scalability, with code publicly archived.","tokens_in":17337,"tokens_out":2708,"duration_ms":31131,"significance":"If correct, the approach would provide a scalable, dynamics-agnostic method for network cluster detection tied to synchronization-relevant symmetries. The release of Fortran code and the application to networks with up to ~146k nodes are concrete strengths, and the synthetic benchmark gives an instructive demonstration. However, the central claims rest on several assumptions that are either internally inconsistent or insufficiently justified, most notably the noisy variant's clustering threshold and the interpretation of MPA magnetizations as synchronization. The paper's value would improve substantially if these gaps were addressed.","major_comments":[{"comment":"There is an internal contradiction in the noisy version. The smoothing threshold epsilon in Eq. (18) is set to machine precision (10^-15 or 10^-18), while the random fields h_i are drawn from a continuous uniform distribution. For generic h_i, the local magnetizations from Eq. (10) will be distinct for every node, so Eq. (18) should give Q=N and p=0 for all J. The non-trivial plateaus in Figs. 4–6 therefore require an undocumented, larger tolerance. Because the WordNet application and the plateau conclusions rely entirely on this variant, the reported results cannot be reproduced from the stated algorithm. Please specify the actual clustering tolerance used, justify why it is not arbitrary, and show the sensitivity of the plateaus to that tolerance.","section":"Sec. VI.C, Eq. (18); Figs. 4–6"},{"comment":"The paper's central claim equates clusters of equal MPA magnetizations with 'clusters of synchronized nodes,' but Sec. VII.A explicitly states that the MPA dynamics is 'an artificial probe and bears no relation to the actual physical dynamics.' The target definition in Eq. (3) uses actual state variables x_i, whereas the algorithm operates on the mean-field magnetizations <x_i>_t of a binary model. No argument or evidence is provided that equality of these thermal averages corresponds to synchronization in any continuous-time oscillator system, especially given the acknowledged low-resolution binary reduction. Please either provide a concrete mapping or substantially soften the synchronization interpretation.","section":"Sec. VII.A; Sec. III, Eqs. (3) and (10)"},{"comment":"The recovery of TE groups in the noiseless positive-initial-condition variant is to a large extent a built-in consequence of the Ising model and the MPA respecting graph automorphisms: uniform couplings and symmetric initial conditions force identical messages and magnetizations for topologically equivalent nodes. Thus Figs. 1–2 confirm this symmetry, but they do not independently validate that the clusters correspond to synchronization under unspecified dynamics. This circularity weakens the paper's main evidence. Please provide a validation against a genuine oscillator dynamics (e.g., Kuramoto or power-grid swing) or clearly frame the method as detecting symmetry-based structures rather than synchronization.","section":"Sec. VI.A, Figs. 1–2"},{"comment":"The algorithm uses local magnetizations at a finite tmax even when the global stationarity condition in Eq. (17) is not satisfied. For the noisy synthetic case only 10–35% of messages are stationary, and for WordNet (tmax=10^3) the paper reports no stationarity fraction at all. If the MPA has not converged, the magnetizations in Eq. (10) are time-dependent, so the resulting partition depends on an arbitrary stopping time. Please report the stationarity fraction for WordNet and demonstrate that the reported Q(J) and p(J) are robust to increasing tmax.","section":"Sec. VI, Eq. (17); Sec. VII.B"}],"minor_comments":[{"comment":"Typo: 'desyncrhronization' should be 'desynchronization'.","section":"Abstract"},{"comment":"The phrase 'with a lager value of tmax' should read 'a larger value.'","section":"Sec. VI"},{"comment":"The label 'Qynch' in the top plot appears to be a typo for 'Qsynch'.","section":"Fig. 3"},{"comment":"The output table shows two nodes both labeled 24 (node 24 appears twice); this is likely a typo and should be corrected for clarity.","section":"Sec. VI.D, Fig. 9"},{"comment":"The claim that the method is 'unbiased' is too strong: Eq. (7) fixes the interaction form to Ising-type couplings, which is already a modeling choice. Please qualify.","section":"Sec. II"}],"recommendation":"major_revision","confidential_remarks":"The noisy-variant inconsistency is the most serious technical issue; I would ask the authors to provide the actual clustering tolerance and rerun the noisy experiments with a clearly specified, physically motivated threshold. The broader synchronization interpretation also needs substantial reworking or stronger validation. The paper has merit as a scalable structural-clustering tool, but in its current form the central claims outrun the evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper's noiseless message-passing method is a reasonable, scalable way to find clusters of nodes with equal Ising magnetizations, and the code on Zenodo is a real plus. But the noisy variant, which drives the WordNet application and the plateau conclusions, has a technical contradiction that needs fixing before the results can be trusted. You set the clustering tolerance to machine precision (Sec. VI, Eq. 18), yet draw random fields from a continuous distribution; generic fields make local magnetizations distinct, so Q should equal N and p=0 for every J. The plateaus in Figs. 4–6 therefore require a larger, undisclosed tolerance. That is a load-bearing gap, not a nitpick.\n\nWhat is genuinely new: the combination of belief propagation on an Ising model with topological-equivalence (TE) groups, pushed as a synchronization probe rather than community detection. The toy example in Fig. 9 is helpful, and the scaling on WordNet (145k nodes in seconds per coupling) is credible. I also credit the paper for explicitly calling the MPA an artificial probe in Sec. VII.A, even if the abstract and intro oversell the connection to real oscillator dynamics.\n\nThe soft spots, in proportion. First, the epsilon inconsistency above. Second, the noiseless positive-initial-condition recovery of TE groups is close to tautological: uniform couplings and symmetric initial conditions force identical messages for automorphism-equivalent nodes. That is a sanity check, not evidence of predictive power. Third, the \"unbiased\" claim is overstated; binary coarse-graining, uniform couplings, and the MPA update rule are all modeling choices. Fourth, the abrupt desynchronization result is a property of the effective Ising model, not of continuous-time synchronization, and the paper should say so much more carefully. Fifth, the plateau universality across h* is asserted from simulations but lacks a mechanistic explanation; given the epsilon issue, those plateaus may be artifacts of an arbitrary clustering threshold.\n\nWho gets value from this: researchers working on network symmetry, cluster synchronization, or fast heuristics for partitioning large graphs. The noiseless algorithm could be useful as a symmetry-based cluster finder. The noisy variant, as written, cannot be accepted.\n\nRecommendation: Yes, send to peer review. The framework is salvageable and the code exists, but the authors must disclose the clustering tolerance actually used, resolve the contradiction with machine-precision epsilon, and substantially temper the synchronization claims unless they add validation against a real oscillator dynamics. A serious referee would push exactly on these points.","headline":"A clean noiseless symmetry-based partitioner with an internally inconsistent noisy variant and an overextended synchronization interpretation.","tokens_in":17778,"tokens_out":4060,"would_cite":false,"duration_ms":44462,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.45.Xt","89.75.Hc"],"model":"deepseek-v4-flash","headline":"A message-passing algorithm on a binary Ising surrogate can partition large networks into clusters of nodes that would synchronize, without knowing the microscopic dynamics.","keywords":["message-passing algorithm","belief propagation","cluster synchronization","network partitioning","topological equivalence","Ising model","random-field Ising model","complex networks"],"falsifier":"On a graph with known TE groups, simulate a concrete continuous-time oscillator dynamics (e.g., Kuramoto with heterogeneous natural frequencies) and compare its synchronized clusters with the MPA clusters for the same coupling strength. If nodes that the MPA puts in one cluster systematically fail to synchronize, or if nodes in different MPA clusters do synchronize, the surrogate claim is refuted. A simpler search: find any graph where the MPA predicts a cluster partition that no plausible continuous dynamics reproduces at any coupling.","tokens_in":16861,"feed_emoji":"🕸️","tokens_out":6729,"duration_ms":71402,"temperature":0.7,"pith_summary":"This paper tries to establish that the synchronization structure of a network can be read off from a surrogate model rather than from any specific oscillator dynamics. The surrogate is a system of binary Ising spins on the same graph with uniform couplings, evolved by the message-passing algorithm; nodes whose stationary local magnetizations coincide are declared a synchronized cluster. The author argues that topologically equivalent nodes (same degree and same external neighborhood) act as nucleation centers, and that the critical couplings where the zero-magnetization state destabilizes organize the cluster hierarchy. Because message passing scales near-linearly, the method is claimed to be usable on networks with hundreds of thousands of nodes, demonstrated on the US power grid and WordNet. A central reported finding is that abrupt desynchronization appears in simple graphs without higher-order interactions, while added noise smooths the transitions into plateaus at larger coupling.","feed_headline":"Message passing finds synchronized clusters in giant networks","feed_subtitle":"A binary spin model run on the graph itself predicts which nodes would sync, without assuming any specific oscillator dynamics.","key_machinery":"The machine is the message-passing (belief-propagation) update for binary variables: each directed edge carries a message ui→j, updated via tanh of the sum of incoming messages (Eq. 9), and the local magnetization of node i is tanh of the total incoming field (Eq. 10). With all couplings equal to J, the trivial zero-message solution loses stability at a critical Jc, and the sequence of critical couplings {J1, J2, ...} gives the cluster-splitting events. Topologically equivalent nodes — groups with identical degree and identical external neighborhoods — are the nucleation sites, and in the noisy variant frozen random fields hi destroy the symmetric structure and replace sharp transitions with","core_discovery":"The central claim is that equal-time conditional probabilities of an unknown stochastic dynamics on a graph can be replaced, without bias, by a binary Ising model whose couplings are solved by message passing. Sweeping the uniform coupling J across the critical points J1, J2, ... of this effective model, the local magnetizations split into groups; nodes with equal magnetization (within machine precision) form the synchronized partitions. Topologically equivalent (TE) groups are shown to nucleate clusters, and clusters tend to absorb entire TE groups as J grows. In the noiseless version with random initial conditions, Q and Qsynch grow and then drop discontinuously at each critical point, yie","pith_inferences":["A testable refinement would be to run known oscillator dynamics on the same graphs and check whether the MPA clusters are the maximal synchronizing sets; the paper only asserts containment for the power grid, not equality.","The binary coarse graining merges distinct TE groups into large clusters; extending the surrogate to multi-state or continuous variables could resolve sub-clusters that the binary version cannot see.","The noise-amplitude-independent plateaus suggest the noisy MPA could serve as a universal descriptor of a network's synchronization propensity, comparable across networks of different sizes and topologies.","If the surrogate hypothesis holds, the same machinery could be adapted to flag consensus-prone or oscillation-prone modules before any dynamics is simulated, as a preprocessing step for control or prediction studies."],"forward_implications":["The approach can partition networks of hundreds of thousands of nodes at near-linear cost, where spectral synchronization methods become infeasible.","No specific dynamical model (Laplacian or Kuramoto) has to be chosen, so the detected clusters are unbiased with respect to the unknown dynamics.","Abrupt desynchronization does not require higher-order interactions; it can arise from ordinary pair couplings in a simple graph.","In the noisy version, the number and height of plateaus in the participation ratio are independent of noise amplitude, giving a stable structural fingerprint.","Topologically equivalent groups are expected to be contained in synchronized clusters, linking structural symmetry to cluster synchronization."],"fun_headline_variants":["Unbiased message passing finds sync clusters at scale","Message passing maps sync clusters in large networks","Scalable message passing finds synchronization clusters","Message passing finds sync clusters without assuming dynamics","Unbiased scalable sync clustering via message passing"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that equal stationary magnetizations of a binary message-passing Ising model stand in for real synchronization, even though the paper states that this surrogate is an artificial probe with no relation to the actual physical dynamics (as for the power grid) and that the binary reduction is a low-resolution coarse graining.","fun_headline_variants_meta":{"raw":{"variants":["Unbiased message passing finds sync clusters at scale","Message passing maps sync clusters in large networks","Scalable message passing finds synchronization clusters","Message passing finds sync clusters without assuming dynamics","Unbiased scalable sync clustering via message passing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000697,"raw_usage":{"total_tokens":2950,"prompt_tokens":671,"completion_tokens":2279,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":415,"completion_tokens_details":{"reasoning_tokens":2213}},"tokens_in":415,"tokens_out":2279,"duration_ms":17241,"temperature":1.0,"reasoning_tokens":2213,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T08:42:39.052498+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a graph with known TE groups, simulate a concrete continuous-time oscillator dynamics (e.g., Kuramoto with heterogeneous natural frequencies) and compare its synchronized clusters with the MPA clusters for the same coupling strength. If nodes that the MPA puts in one cluster systematically fail to synchronize, or if nodes in different MPA clusters do synchronize, the surrogate claim is refuted. A simpler search: find any graph where the MPA predicts a cluster partition that no plausible continuous dynamics reproduces at any coupling.","supporting_citations":[],"review_version":1}