{"id":"a0180dc4-8e75-434f-ba59-fcc666038fd1","arxiv_id":"2412.18322","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":2.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of Graph Mamba, the adaptation of state-space models (Mamba, S4, S6) to graph learning, synthesizing roughly 30 recent papers into a taxonomy of architectures, applications, benchmarks, and open challenges.","lead":"This survey maps the emerging Graph Mamba research area, covering architecture, variants, applications, benchmarks, and open problems for state-space models adapted to graph learning. It is a useful entry point to a fast-moving subfield, but its 'first comprehensive survey' claim, transcribed comparison tables, and URL placeholders need independent verification.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'first comprehensive Graph Mamba survey' claim is never tested; if a comparable survey predates Dec 2024, the flagship novelty claim fails.","rationale":"I read the paper as an attempt to provide the field's first organizing reference for Graph Mamba, with the strongest claim being the Abstract's 'first comprehensive study devoted to Graph Mamba.' For that claim to hold, no comparable prior survey can exist. The paper's Related Surveys section reviews GNN surveys and Mamba surveys but never checks for prior Graph-Mamba surveys, so the decisive premise is asserted rather than demonstrated. This is the same weakest assumption the reader identified, and it is the most load-bearing because it directly supports the novelty and positioning of the entire survey. The other issues the reader listed are concrete and should be fixed, but they are localized transcription and attribution errors rather than threats to the central claim. The proposed search is a single, bounded check: if a prior comprehensive survey surfaces, the flagship claim is falsified; if not, the survey's novelty claim is substantially supported. Since the reader already issued CONDITIONAL, my stress-test does not move the verdict; I recommend UNCHANGED, with the explicit condition that the authors verify the 'first' claim or qualify its scope. I am not asserting that a prior survey exists, only that the paper gives no evidence either way, and the cost of testing is low enough to be a reasonable condition for acceptance.","tokens_in":34073,"tokens_out":4330,"duration_ms":43398,"concrete_test":"Perform a reproducible literature search on arXiv, DBLP, Semantic Scholar, and Google Scholar with no lower date limit and an upper bound of 2024-12-24, using queries that combine ('Graph Mamba' OR 'Mamba' AND 'graph') with ('survey' OR 'review' OR 'overview') and also ('state space model' AND 'graph' AND 'survey'). Screen results against the inclusion criteria implied by the paper's own taxonomy: coverage of architecture, graph types, scanning mechanisms, applications, and benchmarks. If any prior comprehensive survey satisfies these criteria, the 'first' claim fails and must be qualified; if no such survey is found, the claim is supported and the conditional can be lifted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The Abstract asserts 'This survey represents the first comprehensive study devoted to Graph Mamba,' and Section 1.2 repeats that 'there is currently no thorough review.' The whole contribution is premised on this novelty claim. Yet the paper never substantiates it: Section 1.1 surveys only GNN surveys and Mamba surveys, and no search, database, date range, or comparison against prior Graph-Mamba or graph-SSM surveys is reported. Because 'first' is a comparative factual claim, the absence of any evidence for it is a genuine gap in the argument, not just a rhetorical weakness. If a comparable survey on Graph Mamba or state-space models for graph learning existed before 24 December 2024, the central 'first comprehensive study' claim is false, and the paper's organizing-reference value is weakened even though its internal per-domain summaries may remain useful. This is the most load-bearing unverified premise. The concrete transcription issues (URL1-URL15 placeholders in Table 9, the '069' complexity entry in Table 12, and the GAT attribution error in Section 2.1.2) are real but fixable; they do not by themselves overturn the central claim as directly as a prior competing survey would.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript is a survey of Graph Mamba and, more broadly, state-space models (SSMs) for graph learning. It opens with preliminaries on GNNs, SSMs, and the Mamba architecture; proposes a taxonomy of graph types (dynamic, heterogeneous, spatio-temporal) and six selective-scanning mechanisms; reviews applications in general-purpose graph learning, knowledge/heterogeneous graphs, traffic/environmental forecasting, healthcare and biosignals, remote sensing, finance, and sentiment analysis; lists evaluation metrics and datasets; and provides comparative tables of Graph Mamba variants against baseline models in several domains. The paper closes with challenges and future research directions. Its stated flagship contribution is being the first comprehensive survey devoted to Graph Mamba.","tokens_in":34276,"tokens_out":4140,"duration_ms":39157,"significance":"If its coverage and tables are made reliable, this survey could serve as a useful entry point for researchers entering the area: it organizes a fast-growing literature under an architecture-oriented taxonomy, catalogs scanning mechanisms, summarizes training strategies, and gathers applications and benchmark datasets in one place. The comparative tables (Section 6) and the open-problems discussion (Sections 7 and 8) are the most valuable artifacts. The survey does not claim new derivations, code, or experiments, so its significance rests on accuracy and completeness of the synthesized material. The claim of being the first comprehensive Graph Mamba survey is potentially significant but is not substantiated; several concrete transcription errors currently undercut the paper's usefulness as a benchmark reference.","major_comments":[{"comment":"The Abstract and Section 1.2 assert that this is 'the first comprehensive study devoted to Graph Mamba' and that 'there is currently no thorough review' bringing the findings together. Section 1.1 positions the work only against GNN surveys and Mamba surveys; no literature search, database, date range, or comparison with any prior survey on graph state-space models or Graph Mamba is reported. Because 'first' is a comparative factual claim, this is load-bearing for the stated contribution. The authors should either report a documented search (e.g., arXiv, Google Scholar, DBLP with date range and query terms) and discuss any existing graph-SSM surveys, or soften the claim to match what is actually demonstrated.","section":"Abstract and Section 1.2"},{"comment":"The comparative tables, which are a central promised contribution, contain verifiable transcription errors: Table 12 lists GraphMamba's complexity as '069' instead of '0.69'; Table 9 uses 'URL1' through 'URL15' placeholders instead of actual dataset links; and Section 2.1.2 attributes GAT to reference [13] (Hamilton et al., GraphSAGE) while the GAT paper is reference [58]. These errors suggest that the table entries have not been systematically checked against the primary sources. Every row of Tables 10-14 and the dataset table should be re-verified against the original papers, and the citations in Section 2.1.2 should be corrected.","section":"Section 6, Tables 9 and 12; Section 2.1.2"},{"comment":"The paper never states the inclusion criteria that distinguish 'Graph Mamba' from the broader class of 'state-space models for graph learning,' although the title uses both. Section 3.1 describes Graph Mamba as based on state-space message passing, selective scanning, and spatial-temporal integration, but Section 4.4 includes GraphS4mer [47] and other S4-based models that do not use the Mamba selective-scanning architecture. This boundary matters because the coverage claims and the 'first comprehensive Graph Mamba survey' statement depend on a clear definition. Please state the selection criteria used to include or exclude models in each section.","section":"Sections 2.3, 3.1, and 4.4"},{"comment":"The comparative analysis reports numerical results from different primary papers without stating the provenance of the baseline numbers or whether the same experimental protocols, backbones, and hyperparameter settings were used. For example, Table 10 compares STG-Mamba with LSTM-based methods, STGCN, and STAEformer, but the table does not cite the source of each reported value, and the text concludes that the Graph Mamba model is 'superior' without describing how the comparison was made consistent. The authors should provide a citation for each reported baseline, state whether numbers were taken from the original papers or recomputed, and flag any comparisons that mix results from different experimental setups.","section":"Section 6, Tables 10-14"}],"minor_comments":[{"comment":"There are numerous typographical errors, including 'Nnumerous', 'MAmba', 'Tranformers', 'Grah-Mamba', 'Lincan el al.', and 'el al.' in several table entries. A thorough proofreading pass is needed.","section":"Throughout"},{"comment":"The sentence describing fMRI data reads 'brain regions of interest (ROIs.' and is missing the closing parenthesis; this should be corrected.","section":"Section 5.2"},{"comment":"The text defines 'GPUs' as 'Graph Processing Units'; the standard expansion is 'Graphics Processing Units,' and the intended meaning in context should be clarified or corrected.","section":"Section 7.1.2"},{"comment":"The phrase 'Thi will improve the predictive accuracy' should read 'This will improve the predictive accuracy.'","section":"Section 8.3"},{"comment":"In Equation (1), the normalization factor c_ij is introduced as 'a normalization factor' but its definition in terms of node degrees is not given; adding the standard definition would improve readability.","section":"Section 2.1.1"},{"comment":"The column header 'second/epoch' mixes a full word with a slash; use 'seconds/epoch' or a consistent unit format.","section":"Table 13"}],"recommendation":"major_revision","confidential_remarks":"The reference list contains several self-citations in contexts where standard references exist (e.g., refs. [14], [35], [44], [45], [62]); this may warrant editorial attention. The manuscript is formatted as an ACM journal submission with a placeholder DOI, so the venue fit and completeness of the submission should be confirmed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis survey is exactly what it looks like: a clearly organized entry point to the Graph Mamba literature. It brings together the 2024 papers on state-space models for graphs, proposes a reasonable taxonomy of graph types (dynamic, heterogeneous, spatio-temporal) and scanning mechanisms, and summarizes applications from traffic to healthcare. If you're new to the area, this gives you a quick map. Credit where due: the per-domain summaries are largely faithful to the primary papers, and the comparison tables are useful transcriptions.\n\nThe soft spots are real but fixable. First, the flagship claim of being 'the first comprehensive study devoted to Graph Mamba' is asserted, not demonstrated. The related-work section reviews GNN surveys and Mamba surveys but never searches for prior Graph Mamba or graph-SSM surveys. If one exists, the central novelty claim fails. This needs to be checked and, if necessary, qualified. Second, there are concrete transcription errors: GAT is misattributed to the GraphSAGE paper (ref [13]) instead of Veličković et al. (ref [58]); Table 12 lists GraphMamba's complexity as '069' instead of 0.69; Table 9 has placeholder URLs (URL1–URL15). These undermine confidence in the tables as a benchmark reference. Also, the introduction promises 'empirical studies' that the survey never performs; that sentence should go.\n\nNone of this is fatal. The survey derives nothing, so there is no circularity problem in the technical sense. The taxonomy and summaries have value even if the 'first' claim is softened. But the paper needs a heavy revision pass: verify the novelty claim against the actual literature, fix the transcription errors, and remove the promise of empirical evaluation.\n\nWho should read it? Newcomers and researchers looking for a quick map of Graph Mamba applications. Experts won't learn much beyond the organizing frame. It deserves a serious referee because a good survey of this fast-moving subfield is valuable, but I would not accept it in its current form.","headline":"A useful but sloppy survey of Graph Mamba; the 'first comprehensive survey' claim is unverified and the transcription errors need fixing before it can serve as a reliable benchmark reference.","tokens_in":34865,"tokens_out":2354,"would_cite":false,"duration_ms":22466,"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":"This survey claims to be the first comprehensive study of Graph Mamba, organizing the state-space-model approach to graph learning.","keywords":["State Space Models","Mamba Block","Graph Mamba","Graph Learning","Graph Convolutional Network","Selective scan","Spatio-temporal graphs","Heterogeneous graphs"],"falsifier":"A literature search would falsify the 'first comprehensive study' claim if it finds an earlier Graph Mamba survey; a check of Tables 10 to 14 against the cited primary papers would falsify the benchmark claim if entries such as Table 12's '069' complexity figure or the attributed GAT results do not match the originals.","tokens_in":33837,"feed_emoji":"🌐","tokens_out":4892,"duration_ms":42887,"temperature":0.7,"pith_summary":"The paper tries to establish that Graph Mamba, a family of models combining state-space models with graph-structured data, has matured into a distinct research area deserving its own survey, and that this survey is the first to map it. It argues that Graph Mamba addresses central GNN weaknesses, limited long-range dependency capture and quadratic complexity, by using selective scanning mechanisms that process graphs in linear time. If the survey's organization is sound, it gives newcomers and experts a single starting point for the architecture, its variants, applications in traffic, healthcare, finance, remote sensing, and sentiment analysis, plus benchmarks and open problems.","feed_headline":"First comprehensive survey maps Graph Mamba models","feed_subtitle":"One reference for architectures, scanning methods, benchmarks, and open problems across Graph Mamba applications.","key_machinery":"The central machinery is the selective scan mechanism inherited from Mamba and the S6 state-space layer: an input-dependent gating of the state-transition matrices $B$, $C$, and $\\Delta$, followed by discretization and a linear-time state-space evolution. In Graph Mamba this mechanism is adapted to graphs by treating nodes and edges as sequence elements and deciding which states propagate across the structure. The paper organizes the field around this machinery: state-space-based message passing replaces neighborhood aggregation, graph types determine how the scan is applied, and the scanning strategies (graph selective, temporal dependency, directed, bi-directional, recurrent, parallel) specify how traversal is ordered. The claim that Graph Mamba forms a unified area rests on these models sharing this core mechanism while differing in how they linearize graph structure.","core_discovery":"The paper's central claim is that Graph Mamba is a coherent and rapidly growing family of models at the intersection of selective state-space models (S6) and graph learning. Its architecture replaces or augments GNN message passing with state-space-based propagation driven by a selective scan that focuses on the most relevant nodes and edges, enabling long-range spatial and temporal dependency modeling at linear complexity. The survey maps this area through a taxonomy of graph types, dynamic, heterogeneous, and spatio-temporal, along with six scanning mechanisms, graph selective, temporal dependency, directed, bi-directional, recurrent, and parallel, and training strategies. Comparative tables position Graph Mamba variants ahead of LSTM, GCN, attention, and transformer baselines in traffic forecasting, brain classification, hyperspectral image classification, financial prediction, and aspect-based sentiment analysis.","pith_inferences":["If the survey's claim to be first is correct, it may become the organizing reference for this subarea, and its taxonomy of graph types and scanning mechanisms could become a standard vocabulary for describing Graph Mamba models.","The comparative tables suggest that selective state-space models may serve as a drop-in substitute for attention in graph tasks where long-range dependencies and linear scaling matter, but the reported gains come from domain-specific adaptations rather than a single universal architecture.","A testable extension is a shared benchmark suite that evaluates Graph Mamba variants on the same dynamic, heterogeneous, and spatio-temporal graphs with unified metrics, since the tables compile numbers from papers using different splits and protocols.","The repeated emphasis on self-supervised learning as a future direction implies that labeled-graph scarcity, not architecture capacity, may be the next constraint on Graph Mamba adoption."],"forward_implications":["Graph Mamba models can capture long-range dependencies in graphs with reported linear-time complexity, making them a practical alternative to attention-based GNNs on large dynamic graphs.","The taxonomy of scanning mechanisms gives practitioners a menu for adapting state-space models to a given graph type and task, from directed citation networks to bidirectional spatial scans.","The benchmark tables position Graph Mamba variants ahead of LSTM, GCN, and transformer baselines in traffic forecasting, brain-activity classification, hyperspectral imaging, stock prediction, and sentiment analysis.","The survey's list of open challenges, scalability, interpretability, slow convergence on heterogeneous graphs, and training on temporal graphs, defines a concrete research agenda for the field."],"supporting_citations":[{"why":"Supplies the base Mamba selective state-space architecture that Graph Mamba variants adapt to graph-structured data.","marker":"[15]"},{"why":"Introduces the Graph Mamba Network, the core selective state-space model for heterogeneous and dynamic graphs.","marker":"[32]"},{"why":"Presents Graph-Mamba with node prioritization for long-range graph sequence modeling, a key surveyed variant.","marker":"[3]"},{"why":"Defines STG-Mamba, the spatial-temporal graph learning model used as the main traffic forecasting comparison.","marker":"[20]"},{"why":"Offers SpoT-Mamba, with multi-way walk sequences and Mamba-based temporal scans for spatio-temporal forecasting.","marker":"[25]"},{"why":"Introduces HeteGraph-Mamba, a selective state-space model for heterogeneous graphs with graph-to-sequence conversion.","marker":"[24]"},{"why":"Presents BrainMamba, the brain-activity encoding model that anchors the healthcare and biosignal applications and benchmarks.","marker":"[33]"},{"why":"Provides GraphS4mer, a structured-state-space biosignal model that serves as a comparative baseline in brain classification.","marker":"[47]"},{"why":"Proposes GSSC, a graph state space convolution approach that underpins the general-purpose benchmarking discussion.","marker":"[50]"},{"why":"Introduces DYGMAMBA, a continuous-time dynamic graph state-space model used for dynamic link prediction benchmarks.","marker":"[21]"}],"fun_headline_variants":["Graph Mamba: first survey of state-space models for graphs","State-space models for graphs: the Graph Mamba survey","Graph Mamba explained: comprehensive survey of SSM graphs","Graph Mamba: the state-space model guide for graph learning","First comprehensive survey of Graph Mamba state-space models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's value rests on the premise that no comparable Graph Mamba survey existed when it was written, and that the numbers in its comparison tables faithfully reproduce the original papers' results.","fun_headline_variants_meta":{"raw":{"variants":["Graph Mamba: first survey of state-space models for graphs","State-space models for graphs: the Graph Mamba survey","Graph Mamba explained: comprehensive survey of SSM graphs","Graph Mamba: the state-space model guide for graph learning","First comprehensive survey of Graph Mamba state-space models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000516,"raw_usage":{"total_tokens":2484,"prompt_tokens":903,"completion_tokens":1581,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":519,"completion_tokens_details":{"reasoning_tokens":1499}},"tokens_in":519,"tokens_out":1581,"duration_ms":9624,"temperature":1.0,"reasoning_tokens":1499,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:48:07.043463+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A literature search would falsify the 'first comprehensive study' claim if it finds an earlier Graph Mamba survey; a check of Tables 10 to 14 against the cited primary papers would falsify the benchmark claim if entries such as Table 12's '069' complexity figure or the attributed GAT results do not match the originals.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides GraphS4mer, a structured-state-space biosignal model that serves as a comparative baseline in brain classification."}],"review_version":1}