{"id":"88e2bdbd-5730-451e-b82a-cdf7b09075e3","arxiv_id":"2505.12910","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SourceDetMamba detects rumor sources by feeding reverse-ordered hypergraph snapshots through a Mamba state-space model with a graph-aware state update, and reports large accuracy gains over baselines on eight datasets.","lead":"The paper presents SourceDetMamba, a state-space model that takes a series of hypergraph snapshots of an unfolding rumor and scans them in reverse to identify the rumor's origin nodes. If it holds up, it offers a faster and more accurate way to trace misinformation back to its starting points in group-based social networks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SourceDetMamba receives exact infection timestamps (Eq. 4) while baselines are not shown to; a timestamp threshold may already solve the synthetic task, so the SOTA margin is unverified.","rationale":"The reader's weakest assumption was that the synthetic cascade generator may not transfer to real dynamics; that is a valid external-validity concern. My stress test focuses on an internal comparison issue that is more immediately decisive: the task may be solvable by reading the timestamp feature, and baselines are not described as having access to it. If the timestamp-threshold diagnostic succeeds, the stated 10%-23% gains are not evidence about the graph-aware state update mechanism, regardless of whether the simulation is realistic. The paper deserves credit for the ablation study (Table 2), which supports the role of PEs, sequential snapshots, and graph communication; however, the omission of a timestamp ablation and the lack of any stated baseline input protocol leave the central SOTA claim conditional. I therefore keep the reader's CONDITIONAL verdict rather than moving to reject: the concern is concrete but testable, and a fair-input rerun could resolve it. No code, seeds, or error bars are provided, which compounds the uncertainty but is not the load-bearing issue itself.","tokens_in":13731,"tokens_out":8780,"duration_ms":97430,"concrete_test":"Add a 'w/o timestamps' condition by setting X2_i = -1 for all nodes in SourceDetMamba and rerunning Table 1. Also run a trivial baseline on the same synthetic test split that predicts the source set as the informed nodes with minimal t_{v_i} (or t=0 where available). If the trivial timestamp rule reaches an F-Score close to SourceDetMamba's, or if removing X2_i drops F-Score materially, the claimed improvement over state-of-the-art methods is not a fair test of the proposed architecture and the comparison must be rerun with baselines given identical snapshot and timestamp inputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 gives SourceDetMamba the feature X2_i = t_{v_i} for every informed node (Eq. 4). In the Section 5.1 evaluation, propagation starts with 5% randomly chosen sources, so under the heterogeneous IC model the true sources are exactly the nodes whose infection time is t=0. A prediction rule that flags informed nodes with t=0 would therefore identify the source set directly. The paper never states that the nine baselines receive this timestamp feature; Section 5.1 only says the hypergraph snapshots are clique-expanded before being applied to baselines. Because the baseline descriptions in Section 2.1 rely on node states or fitted diffusion models rather than recorded infection times, the 10%-23% accuracy advantage claimed in Section 5.2 may stem from an input-information asymmetry rather than from the proposed graph-aware Mamba mechanism. Table 2 ablates positional encodings and the graph-aware state update but never ablates the timestamp feature, so the contribution of X2_i to the reported F-Scores is unknown. This is the weakest point in the empirical argument for the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SourceDetMamba, a state-space-model-based method for source detection in sequential hypergraph snapshots. Hypergraph snapshots are embedded with user states, diffusion timestamps, and Laplacian positional encodings, then fed in reverse chronological order into Mamba blocks. A graph-aware state update mechanism propagates node states through hyperedges, and the final sequence element is used for per-node source classification. Experiments on eight datasets with synthetic hypergraph propagation report consistent and substantial improvements over nine baselines in ACC, F-Score, and AUC, plus an ablation study of the main components.","tokens_in":13949,"tokens_out":5113,"duration_ms":53959,"significance":"If the empirical claims hold, the paper provides a meaningful step toward using linear-complexity sequence models for temporal source localization on higher-order networks. The idea of reverse-order snapshot sequences and a graph-aware state update is interesting and the ablations are largely consistent. However, the evaluation rests on a potentially privileged input feature (exact infection timestamps), and the results are reported without variance or code release, so the magnitude of the stated advantage over baselines is not yet verified. The paper also contains a technical inconsistency in its description of the state-space convolution kernel.","major_comments":[{"comment":"The diffusion timestamp feature X2_i = t_{v_i} is given only to SourceDetMamba for every informed node. In the Section 5.1 simulation, propagation starts at t=0 with 5% randomly selected sources, so the true sources are exactly the informed nodes with t_{v_i}=0. A trivial rule that flags these nodes would solve the task. The paper does not state that the nine baselines receive this timestamp feature; Section 5.1 only mentions that hypergraph snapshots are clique-expanded before being applied to baselines. Table 2 ablates positional encodings and the graph-aware state update, but never ablates the timestamp feature, so the 10%-23% accuracy advantage claimed in Section 5.2 may stem from an input-information asymmetry rather than from the proposed Mamba mechanism. Please either remove the timestamp feature from SourceDetMamba, provide equivalent temporal information to all baselines, or add an ablation 'w/o timestamps' and a comparison against a threshold-based rule (informed and t=0) to establish that the reported gains survive without this privileged information.","section":"§4.1 Eq. (4); §5.1; §5.2"},{"comment":"All evaluations are on synthetic cascades generated by a heterogeneous independent cascade model with 5% random sources and p_i ~ U(0,0.5). The central claim of 'consistently outperforms state-of-the-art approaches' therefore depends on the faithfulness of this simulation to real rumor dynamics on hypergraphs. The paper does not justify that the reversal-learning signal remains strong under different thresholds, recovery, or observation noise. Table 3 tests SIR/SIS but only for SourceDetMamba, not for baselines, and no variance is reported. Please add experiments with observation noise or partial observability, report baseline performance under SIR/SIS, and explicitly discuss this synthetic-data limitation in the text.","section":"§5.1; Table 3"},{"comment":"The paragraph states that 'by utilizing convolution kernel K = (CB, CAB, ..., CA^k B, ...), we achieve the sequence-to-sequence transformation through y = X' * K'. This convolution representation is valid for a linear time-invariant (LTI) state space model, but the paragraphs immediately before introduce Mamba's selective SSM, where B, C, and Δ are input-dependent functions, so a fixed convolution kernel does not exist. The description as written is internally inconsistent. Please correct this to state that the selective-scan (input-dependent) formulation is used, and that the convolution form applies only in the non-selective LTI case, or remove the fixed-K statement entirely.","section":"§4.2"}],"minor_comments":[{"comment":"No standard deviations, confidence intervals, or repeated runs are reported. Since neural network training is stochastic, please report mean and standard deviation over at least five random seeds for the proposed method and all baselines.","section":"§5.1; Table 1"},{"comment":"The eight datasets are standard machine learning benchmark datasets used as hypergraph structures; calling them 'real-world datasets' for rumor source detection is misleading because the cascades are simulated. Please clarify that the evaluation is synthetic propagation over real-world hypergraph structures.","section":"§5.1"},{"comment":"The symbol Ω is overloaded: it denotes the given hyperedge weight matrix in Eqs. (1) and (9) but also the learned hyperedge weights in Eqs. (15)-(16). Please use a different symbol (e.g., Ω̂) for the learned weights.","section":"§3; §4.2"},{"comment":"In 'y = X′ ∗ K', the star operator is not defined; please specify that it denotes temporal convolution along the sequence dimension, to avoid confusion with graph convolution used elsewhere.","section":"§4.2"},{"comment":"The sentence 'for a sample x with corresponding label z, L(x,z) = −log(x) × z' is unclear; please rewrite using standard cross-entropy notation for binary classification.","section":"§4.3"},{"comment":"The robustness experiment with SIR/SIS models reports only SourceDetMamba results. For a fair robustness claim, please also report the performance of at least the strongest baselines under these diffusion models.","section":"§5.3; Table 3"},{"comment":"The runtime comparison appears to be on a single device without timing details (e.g., whether timings include data loading) or error bars. Please specify the exact hardware/software configuration and the number of independent runs.","section":"§5.2; Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's strongest baseline is GIN-SD, which shares authors with this work, and SourceDetMamba outperforms it by 13-25 F-Score points across datasets. Given the timestamp input asymmetry, this margin should be treated with caution until the ablation requested in Major Comment 1 is performed. The paper also releases no code, which raises reproducibility concerns for a purely empirical contribution. The synthetic-only evaluation is acceptable if framed as a controlled study, but the authors should be explicit about its limits."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: SourceDetMamba is a sensible new combination — reverse-order sequential hypergraph snapshots fed into Mamba with a graph-aware state update (Eq. 16) — and the ablations show the design choices matter. But the central numerical claim does not hold up as presented.\n\nThe stress-test note you passed me lands. Section 4.1 gives the model X2_i = t_{v_i}, the exact infection timestamp, for every informed node. Section 5.1 says propagation starts with 5% randomly selected sources at t=0. So the true source set is exactly the set of informed nodes with timestamp 0. A trivial threshold on that feature solves the task or at least gives a very strong prior. The baseline descriptions in Section 2.1 and the clique-expansion note in Section 5.1 never say the baselines receive timestamps; they get user states or fitted diffusion models. If that asymmetry is real, the claimed 10-23% accuracy advantage is not evidence for the proposed graph-aware Mamba mechanism. The paper never ablates the X2 feature, so we cannot tell how much of the gain is from the brute-force timestamp signal.\n\nNone of this kills the architecture. The graph-aware state update with hyperedge aggregation and learned hyperedge weights is a reasonable extension of Mamba to sequential hypergraphs, and the reverse-order input idea is natural. The ablation section is informative: removing node communication or edge weights hurts, which supports the mechanism's role. The runtime comparison is fine as far as it goes.\n\nOther soft spots, in proportion: no code, no seeds, no error bars — every table entry is a single run on synthetic cascades. The convolution-kernel claim (y = X' * K) is not reconciled with the graph-aware update, which adds an extra term; that is a minor presentation gap. The class-balance weight xi in Eq. 17 is standard ground-truth reweighting, not a circularity.\n\nWho is this for? People working on source detection or temporal graph learning might read the methods section for the architecture idea. I would not cite the reported numbers until the evaluation is fixed. The paper deserves a serious referee: the idea is testable, and the authors are likely capable of shipping code. The right outcome is major revision asking for feature-matched baselines, an ablation of X2, and code release.","headline":"The architecture is a plausible new combination, but the empirical claim is undermined by an input-information asymmetry: the model gets infection timestamps and the baselines apparently don't, which could explain the entire reported margin.","tokens_in":14492,"tokens_out":3023,"would_cite":false,"duration_ms":33035,"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":"SourceDetMamba claims that replaying hypergraph rumor snapshots in reverse order through a Mamba state space model with a graph-aware state update detects rumor sources more accurately and faster than state-of-the-art baselines.","keywords":["source detection","hypergraphs","state space models","Mamba","rumor propagation","sequential snapshots","source localization","graph neural networks"],"falsifier":"Train SourceDetMamba on the paper's synthetic cascades and evaluate it on real social-media rumor traces with independently verified ground-truth sources using the same three snapshots; if its F-Score no longer beats GIN-SD and IVGD by the reported margins, the central claim that reverse-order snapshot learning generalizes to real propagation dynamics is falsified.","tokens_in":13519,"feed_emoji":"🕵️","tokens_out":6499,"duration_ms":58540,"temperature":0.7,"pith_summary":"SourceDetMamba claims that rumor sources can be located by replaying the observed propagation backward: feed hypergraph snapshots of an ongoing rumor to a state space model in reverse chronological order, and the model learns the dynamics that led to the observed state. The paper extends the Mamba architecture with a graph-aware state update, so each node's hidden state is refined simultaneously by the temporal sequence and by messages from its hyperedges. On eight datasets with synthetic cascades, the method reports 10–23% higher accuracy than propagation-aware baselines, 25–35% higher than user-state baselines, and about twice the accuracy of source-centrality methods, while keeping runtime low. The value, if the claim holds, is a source-detection approach that avoids fitting a propagation model as a separate error-prone step.","feed_headline":"Rewinding rumor snapshots lifts source-detection accuracy 10–23%","feed_subtitle":"A Mamba model with a topology-aware state update beats state-of-the-art baselines on eight datasets.","key_machinery":"The load-bearing mechanism is the graph-aware state update inside a Mamba state space model. Mamba is a selective state space model that processes sequences in linear time by maintaining a hidden state; here the hidden state of each node is not updated only from the sequence input $B x_t$ and the recurrent term $A h_{t-1}$, but also gets a topological term $H D_E^{-1} \\Omega H^T D_V^{-1} h_{t-1}$ that propagates and refines states through hyperedges in two stages. This term is what lets the model combine temporal dependencies with hypergraph structure, and the learned hyperedge weights $\\Omega$ let different hyperedges contribute according to their role in the spread.","core_discovery":"The central claim is that SourceDetMamba detects rumor sources more accurately than existing methods because it learns the propagation process directly from sequential hypergraph snapshots rather than from a fitted model or a single snapshot. Snapshots are fed in reverse order, so the first element the model sees is the most recent state and the last is the earliest, and the final output element is used to identify sources. To keep the topology from being lost, the paper adds a graph-aware state update: each node's state is updated as $h_t = A h_{t-1} + B x_t + H D_E^{-1} \\Omega H^T D_V^{-1} h_{t-1}$, where the added term aggregates neighbor states through node-to-hyperedge and hyperedge-to-node passes, with hyperedge weights learned from the current state. On eight datasets, the method reports consistently higher ACC, F-Score, and AUC than all baselines, with the largest claimed gains over propagation-aware methods.","pith_inferences":["A testable extension is to train on synthetic cascades and evaluate on real propagation traces with verified ground-truth sources; if the reported gains shrink, the synthetic-cascade assumption would be the bottleneck.","The same reversal-learning recipe could apply to other inverse problems on temporal hypergraphs, such as identifying the first adopters in opinion dynamics or the origin of a computer worm, where reverse-ordered snapshots are available.","The graph-aware state update is a generic plug-in for Mamba-style models, so it could be used not just for source detection but for any sequential task on hypergraphs that needs topology-aware hidden states.","Because hyperedge weights are learned per state, the model may implicitly identify which group interactions carried the rumor, giving a by-product explanation of the main transmission channels—something the paper does not test."],"forward_implications":["On cascade data generated by similar hypergraph contagion models, SourceDetMamba should keep beating propagation-fitting baselines across the eight datasets evaluated, since the multi-snapshot reverse-order signal is the source of the gain.","Removing any component of the design—the sequential snapshots, the positional encodings, or the graph-aware state update—degrades performance, so each part is doing load-bearing work.","The method stays accurate under SI, SIS, SIR, and IC diffusion models, with a slight drop when sources can recover, so the reversal signal is not tied to one specific spreading rule.","Because the model reads only snapshots and outputs source scores directly, it avoids the extra error of first fitting a propagation model to the observed state.","Snapshot interval matters: intervals too small miss dynamics and intervals too large over-weight non-source influence, so choosing an appropriate observation timescale is part of the method's success."],"supporting_citations":[{"why":"Supplies the Mamba selective state space model that the paper adapts with a graph-aware state update.","marker":"[Gu and Dao, 2023]"},{"why":"Supplies the hypergraph neural network aggregation (node-to-hyperedge, hyperedge-to-node) used to preprocess sequential snapshot features.","marker":"[Feng et al., 2019]"},{"why":"GIN-SD is the strongest propagation-feature baseline and the main accuracy competitor the method must beat.","marker":"[Cheng et al., 2024b]"},{"why":"LPSI is the source-centrality baseline whose performance SourceDetMamba reports roughly doubling.","marker":"[Wang et al., 2017]"},{"why":"IVGD is a propagation-fitting baseline that SourceDetMamba claims to beat by 10–23% in accuracy.","marker":"[Wang et al., 2022]"},{"why":"SL-VAE is another variational propagation-model baseline used in the comparison.","marker":"[Ling et al., 2022]"},{"why":"GCNSI is the representative user-state GCN baseline against which the method claims 25–35% higher accuracy.","marker":"[Dong et al., 2019]"},{"why":"Graph Mamba is the prior state space model on graphs that the paper distinguishes by targeting single graphs rather than sequential snapshots.","marker":"[Behrouz and Hashemi, 2024]"}],"fun_headline_variants":["Reverse-time snapshots boost rumor source detection up to 23%","Hypergraph-aware Mamba spots rumor origins with 10-23% gain","Rewinding rumor cascades nails sources, Mamba style","Topology-aware state update sharpens source finding","Sequential hypergraph snapshots rewind rumors to origin"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains depend on cascades generated by the paper's synthetic heterogeneous independent cascade model—5% random sources, $p_i \\sim U(0,0.5)$, high-order thresholds $p_\\Delta = 0.3$ times the informed fraction, snapshots at 10/20/30% informed—faithfully representing real rumor dynamics; if real cascades have different thresholds, recovery, or observation noise, the reversal-learning signal may disappear.","fun_headline_variants_meta":{"raw":{"variants":["Reverse-time snapshots boost rumor source detection up to 23%","Hypergraph-aware Mamba spots rumor origins with 10-23% gain","Rewinding rumor cascades nails sources, Mamba style","Topology-aware state update sharpens source finding","Sequential hypergraph snapshots rewind rumors to origin"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000414,"raw_usage":{"total_tokens":2122,"prompt_tokens":913,"completion_tokens":1209,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":1124}},"tokens_in":529,"tokens_out":1209,"duration_ms":9649,"temperature":1.0,"reasoning_tokens":1124,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:24:06.548782+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SourceDetMamba on the paper's synthetic cascades and evaluate it on real social-media rumor traces with independently verified ground-truth sources using the same three snapshots; if its F-Score no longer beats GIN-SD and IVGD by the reported margins, the central claim that reverse-order snapshot learning generalizes to real propagation dynamics is falsified.","supporting_citations":[{"cited_title":"Hypergraph neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the hypergraph neural network aggregation (node-to-hyperedge, hyperedge-to-node) used to preprocess sequential snapshot features."},{"cited_title":"Multiple source detection without knowing the underlying propagation model","cited_arxiv_id":null,"evidence_quote":"LPSI is the source-centrality baseline whose performance SourceDetMamba reports roughly doubling."},{"cited_title":"An invertible graph diffusion neural network for source localization","cited_arxiv_id":null,"evidence_quote":"IVGD is a propagation-fitting baseline that SourceDetMamba claims to beat by 10–23% in accuracy."},{"cited_title":"Source localization of graph diffusion via variational autoencoders for graph inverse problems","cited_arxiv_id":null,"evidence_quote":"SL-VAE is another variational propagation-model baseline used in the comparison."},{"cited_title":"Multiple ru- mor source detection with graph convolutional networks","cited_arxiv_id":null,"evidence_quote":"GCNSI is the representative user-state GCN baseline against which the method claims 25–35% higher accuracy."}],"review_version":1}