{"id":"4b64266d-9a3a-4508-8806-1cdee4a617ec","arxiv_id":"2509.06609","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A systematic taxonomy of generalization in graph anomaly detection, spanning transfer learning and foundation models, with an up-to-date review and future directions.","lead":"This paper surveys how graph anomaly detection models are being made to generalize, from transfer learning between related datasets to one-for-all foundation models. It organizes the emerging literature into a taxonomy and lists open challenges, giving researchers a structured map of the field.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The survey's central 'comprehensive review' claim rests on an undocumented paper-selection process; without a reproducible search and explicit inclusion criteria, representativeness and taxonomy assignments cannot be audited.","rationale":"The reader's conditional verdict identifies the absence of a formal selection protocol as the weakest assumption, and this stress-test agrees: the paper's central contribution is a comprehensive map, and comprehensiveness is only meaningful if the corpus is defined and reproducible. The paper is otherwise well-structured, with a plausible taxonomy, honest discussion of open challenges, and useful organization of recent work. The concern is not that the authors are biased or careless, but that the survey's own standards are not applied to its selection. The proposed check would settle the concern by turning an unverifiable claim of comprehensiveness into a measurable comparison against a systematic search. Because the reader already conditioned acceptance on addressable revision, the verdict remains unchanged: CONDITIONAL acceptance until the corpus selection and the classification of borderline methods such as cross-granularity detectors are justified with explicit criteria.","tokens_in":11560,"tokens_out":9040,"duration_ms":85439,"concrete_test":"Reconstruct the survey corpus by executing a defined search on DBLP and arXiv for 2019-2025 with explicit queries combining 'graph anomaly detection' or 'graph outlier detection' with 'transfer learning', 'domain adaptation', 'domain generalization', 'out-of-distribution', 'zero-shot', 'foundation model', or 'generalization'. Apply the paper's Definitions 1-3 as inclusion criteria and compare the resulting qualifying set with the methods covered in Sections III and IV. If a material fraction of qualifying methods is absent (e.g., more than 10-15% of the qualifying set), the comprehensiveness claim is not supported. As a secondary audit, check each Section IV.A method's original evaluation to record whether it uses a test dataset unseen during training, since Definition 3 makes unseen-dataset prediction a defining property of GAD foundation models.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that this paper provides a comprehensive and systematic survey of generalized GAD. For that claim to hold, the set of reviewed methods must be representative of the field, and the taxonomy must reliably partition that set. The manuscript provides no search strategy, no inclusion/exclusion criteria, no database list, no query terms, no date range, and no screening procedure. The paper itself notes in Sec. V that standardized evaluation protocols are lacking for generalized GAD, but it never applies any such protocol to its own corpus selection. Consequently, the boundary between 'transfer learning' and 'foundation models' is enforced narratively rather than by criteria. For example, GOODAT [38] is placed under transfer learning despite not using GAD source data, while cross-granularity multi-task detectors such as HO-GAT [8] and HeagNet [5] are labeled foundation models without evidence of unseen-dataset generalization as required by Definition 3. Without a reproducible corpus, a reader cannot distinguish a systematic map from a curated sample, so the headline claim is not verifiable as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper surveys generalization in graph anomaly detection (GAD), organizing the literature into two paradigms: transfer learning and foundation models. It formalizes conventional GAD, transfer learning, and foundation-model problem settings (Definitions 1-3), proposes a two-level taxonomy (generalization-centric training vs. source-target representation alignment; target-aware pre-training vs. test-time fine-tuning; cross-granularity vs. cross-scenario foundation models), reviews representative methods in each category, and closes with four future directions. The review emphasizes design principles (e.g., data augmentation, representation alignment, prompt/prototype adaptation, feature-space standardization, LLM-based detectors) rather than experimental comparisons.","tokens_in":11775,"tokens_out":6642,"duration_ms":58920,"significance":"The paper addresses a real gap: prior GAD surveys do not focus specifically on generalization. The formal definitions are largely clear, the organization by design principle is intuitive, and the figures convey the paradigm distinctions well. If the corpus is accepted as representative, the taxonomy and problem formalization would help researchers position new work. The paper's strengths include its clear definitions, helpful schematic figures, and an up-to-date reference list; it does not make quantitative predictions or provide code/proofs, so those verification tools are not applicable to a survey. The main caveat is that the survey's central claim of being 'comprehensive' is not verifiable from the manuscript alone, and several taxonomy assignments appear to conflict with the paper's own definitions; these issues are fixable and do not invalidate the underlying organization.","major_comments":[{"comment":"Definition 3 defines a GAD foundation model as a scoring function that predicts anomalies for arbitrary samples from an unseen dataset not in the training collection, with both cross-dataset and cross-granularity capability. However, Section IV-A classifies HO-GAT [8], HeagNet [5], BOURNE [14], UniGAD [13], UniFORM [35], SIGNET [20], GRAM [47], and ASD-HC [36] as cross-granularity GAD foundation models without presenting any evidence that these methods are evaluated on datasets outside their training set; as described, they are multi-task detectors for node-, edge-, and graph-level anomalies on a single graph or domain. This mismatch between the formal definition and the taxonomy means the foundation-model category currently includes methods that do not satisfy the paper's own 'unseen dataset' condition. Please either add per-method evidence of cross-dataset evaluation, tighten the definition, or rename the category (e.g., 'multi-granularity GAD').","section":"II-C and IV-A"},{"comment":"GOODAT [38] is included in the transfer-learning branch, both under generalization-centric training and under test-time fine-tuning. Yet the description states that GOODAT 'repurposes a pre-trained GNN classifier without fine-tuning' and does not mention any source dataset or knowledge transfer from GAD data. This conflicts with Definition 2, which requires 'additional data resources from one or more source datasets.' If GOODAT is meant to be covered, Definition 2 should be broadened to test-time adaptation under distribution shift; otherwise GOODAT belongs to a separate OOD-detection category rather than transfer learning for GAD.","section":"III-A.1 and III-B.2"},{"comment":"The paper's headline claim is that it provides 'a comprehensive and systematic survey of generalization in GAD.' The manuscript does not report a search strategy, database list, query terms, date range, inclusion/exclusion criteria, or screening procedure, so the representativeness of the corpus cannot be audited. This is a load-bearing omission for a survey: without it, the reader cannot distinguish a systematic map from a curated sample, and the taxonomy's completeness is not checkable. Please add a survey-methodology paragraph or appendix that details corpus construction, and consider a table listing each reviewed method with its categorized level and evaluation setting.","section":"Introduction and Section II (overall)"},{"comment":"The boundary between the two main paradigms is not enforced by operational criteria. Several methods classified as foundation models (e.g., ARC [21], AnomalyGFM [31], UNPrompt [27]) are also described as using few-shot target data and in-context learning, which functionally resembles transfer learning, while AdaGraph-T3 appears in both transferable-knowledge learning and test-time fine-tuning without an explicit explanation of the overlap. Since the taxonomy is the paper's main contribution, defining a crisp, testable separation (for example, based on whether training data include multiple domains or whether feature spaces are unaligned) and applying it consistently would strengthen the paper.","section":"II-C and III"}],"minor_comments":[{"comment":"The abstract contains a typo: 'throughtransfer learning' should be 'through transfer learning'.","section":"Abstract"},{"comment":"The label 'Auxiliary modules and constrains' should read 'Auxiliary modules and constraints'.","section":"Figure 2"},{"comment":"The paper alternates between 'foundation models' and 'foundational models' (for example, in the taxonomy paragraph); standardize on one term, preferably 'foundation models'.","section":"II-C"},{"comment":"The subsection title 'Test-Time Fine-Tuning' is misleading for GOODAT [38], which is explicitly described as operating without fine-tuning; consider renaming the subsection to 'Test-Time Adaptation' and noting which methods actually fine-tune.","section":"III-B"},{"comment":"The sentence 'So far, only [1] has explored this issue via heuristic data selection strategies' overstates the contribution of Wild-GAD [1]; the cited paper is a data-selection method rather than a theory of transferability, so the wording should be softened.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The manuscript would benefit from a clear statement of how the reference set was compiled, especially because a substantial share of the surveyed papers involve the corresponding author's group. This is a common pattern in active research areas, but without a search strategy it weakens the independence of the corpus. I do not see grounds for rejection; the taxonomy and definitions can be made consistent with modest revisions, and the survey has a distinct angle relative to existing GAD surveys. The editor may wish to ask the authors to state the literature cutoff date and any deliberate scope exclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the useful part: this is the first survey I've seen that explicitly frames generalization in GAD as a spectrum from transfer learning to foundation models. The problem definitions (Definitions 2 and 3) are clear enough to be usable, and the sub-taxonomy — generalization-centric training vs. source-target alignment on the transfer side, cross-granularity vs. cross-scenario on the foundation side — is a genuinely new way to organize the literature. The sketch figures are helpful, and the discussion of open challenges (theoretical guarantees, evaluation protocols) is honest and on point. If you need a map of this subfield, this is the best one available.\n\nThe soft spots are real but not fatal. The biggest one is the comprehensiveness claim. The paper says \"comprehensive and systematic\" but gives no search strategy, no inclusion/exclusion criteria, no database or query terms. For a survey, that makes the coverage unverifiable. I also noticed a heavy share of self-citations; that alone is not a flaw, but combined with the lack of a selection protocol, it does make you wonder how representative the corpus is.\n\nMore substantive: the category boundaries are porous. AdaGraph-T3 appears in both generalization-centric training and test-time fine-tuning, and GOODAT shows up in two places as well. Cross-listing is fine if the paper says so, but it doesn't. Worse, HO-GAT and HeagNet — multi-granularity detectors trained on a single dataset — are labeled foundation models even though Definition 3 requires generalization to unseen datasets. That looks like a definitional overreach, and it weakens the taxonomy's reliability.\n\nNone of this kills the paper. The taxonomy is still a plausible frame, and the survey is honest about the field's weaknesses. With an added selection protocol and a tightening of the foundation-model label, it becomes a solid reference. As it stands, I'd take the taxonomy with a grain of salt but still use it as an entry point.\n\nVerdict: deserving of serious peer review, but needs revision. I'd want to see the corpus-building process documented before accepting.","headline":"A useful first taxonomy for generalization in GAD, but the comprehensiveness claim outruns the documented selection process.","tokens_in":12195,"tokens_out":2405,"would_cite":true,"duration_ms":21719,"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 all generalized graph anomaly detection methods fit under two paradigms, transfer learning and foundation models, and it supplies a systematic taxonomy to organize them.","keywords":["graph anomaly detection","transfer learning","foundation models","generalization","taxonomy","graph neural networks","zero-shot anomaly detection","cross-domain detection"],"falsifier":"Checking every generalized GAD paper published in the five years before the survey, one could ask whether each falls cleanly into exactly one of the four taxonomy leaves; finding a peer-reviewed method that fits none of them—or that the survey's own categories force into two places at once—would falsify the claim that the taxonomy is comprehensive and systematic.","tokens_in":11397,"feed_emoji":"🧭","tokens_out":4969,"duration_ms":41709,"temperature":0.7,"pith_summary":"This survey sets out to establish that the scattered work on generalization in graph anomaly detection (GAD) can be organized into a single taxonomy with two main branches. The first branch, transfer learning for GAD, reuses knowledge from related source graphs to detect anomalies on a target graph. The second branch, GAD foundation models, aims for a 'one-for-all' model that scores anomalies on unseen graphs, possibly across different granularities such as nodes, edges, subgraphs, and whole graphs. The paper formalizes the problem settings for conventional GAD, transfer learning, and foundation models, and it classifies existing methods into four finer categories. A sympathetic reader would care because a reliable map lets practitioners choose methods by the kind of shift they face and lets researchers see where the design space is still empty.","feed_headline":"Graph anomaly detection gets a taxonomy of generalization","feed_subtitle":"Survey maps methods for shifting graphs and unseen data, from transfer learning to one-for-all foundation models.","key_machinery":"The load-bearing object is the taxonomy itself, built on three formal definitions: conventional GAD (training and testing on one dataset), transfer learning for GAD (using source datasets to improve detection on a target dataset), and GAD foundation models (one model scoring arbitrary samples from unseen datasets). The taxonomy's first branch is split by two challenges—learning transferable knowledge versus capturing target-specific patterns—and the second by two generalization targets—cross-granularity versus cross-scenario. These definitions do the work of the survey: they supply the coordinates in which each reviewed method is placed, and they turn the otherwise anecdotal progress of the field into a map with named design choices such as generalization-centric training, source-target representation alignment, target-aware pre-training, test-time fine-tuning, graph feature standardization, task-agnostic detection, and resource-efficient adaptation.","core_discovery":"The paper's central claim is that generalization in GAD has followed a clear evolution, and that every method along this path can be placed under one of two paradigms. In transfer learning for GAD, the model uses source-domain data to help a target domain under the assumptions that the domains share common knowledge and that their discrepancy is moderate; methods must both learn transferable knowledge and capture target-specific patterns. In GAD foundation models, a single scoring function is trained to accept arbitrary samples from unseen datasets, even from domains and granularities never seen in training; current work splits into cross-granularity models that detect anomalies at several levels at once and cross-scenario models that standardize features, learn task-agnostic detectors, and adapt efficiently. By organizing the literature this way, the authors claim to reveal the design principles behind recent generalized GAD methods and to identify open problems, including the lack of theoretical guarantees on transferability, missing standardized evaluation protocols, the road to universal one-for-all models, and human-in-the-loop adaptation.","pith_inferences":["The taxonomy suggests a natural test for the field: if cross-scenario foundation models keep improving, transfer learning for GAD may become a special case—when the target is known and labeled—rather than a separate paradigm.","Because the survey separates generalization-centric training from test-time fine-tuning, one can read future papers as combinations of these choices; a likely next step is a benchmark that compares methods across all four branches on identical shifts.","The emphasis on feature-space standardization implies that progress on unified graph representations, rather than detector architectures, may be what unlocks broader zero-shot generalization.","A reader should watch whether the categories remain stable as new methods appear: the boundary between learning transferable knowledge and capturing target-specific patterns is defined by target-data availability, which future methods may blur with continual learning."],"forward_implications":["A practitioner facing scarce target data can choose a transfer learning method by whether target graphs are available during training: generalization-centric training when they are not, representation alignment when they are.","Methods that unify node-, edge-, and graph-level detection into one framework (cross-granularity foundation models) imply that correlations across granularities can be exploited rather than treated as separate tasks.","The existence of zero-shot GAD foundation models implies that anomaly detection can be applied to a brand-new graph without retraining, provided features can be standardized into a common space.","The survey's open-challenge list implies that the field's next bottleneck is not more architectures but theory for why transfer works, unified benchmarks, and scalable one-for-all models."],"supporting_citations":[{"why":"Supplies the notion of domain generalization that frames the paper's motivation for moving beyond same-distribution GAD.","marker":"[52]"},{"why":"Establishes the conventional GAD problem and its underlying assumptions, which the survey contrasts with generalized paradigms.","marker":"[25]"},{"why":"Provides a recent broad survey of deep graph anomaly detection that the paper positions itself against as more focused on generalization.","marker":"[32]"},{"why":"Pioneers cross-domain graph anomaly detection and serves as the anchor method for the transfer-learning branch.","marker":"[4]"},{"why":"Exemplifies anomaly-aware one-class domain alignment, a key technique in the source-target representation alignment category.","marker":"[39]"},{"why":"Unifies multi-level graph anomaly detection, representing the cross-granularity foundation model category.","marker":"[13]"},{"why":"Demonstrates zero-shot generalist graph anomaly detection via unified neighborhood prompts, anchoring the cross-scenario foundation model category.","marker":"[27]"},{"why":"Provides a graph foundation model for zero/few-shot anomaly detection, supporting the cross-scenario and resource-efficient adaptation discussion.","marker":"[31]"},{"why":"Introduces a generalist graph anomaly detector with in-context learning, used as a central example of resource-efficient adaptation.","marker":"[21]"},{"why":"Supplies a benchmark for unsupervised outlier node detection that the paper cites when calling for standardized evaluation protocols.","marker":"[16]"}],"fun_headline_variants":["Survey taxonomizes graph anomaly detection generalization","From transfer learning to foundation models: GAD survey","Two paradigms explain GAD generalization","Graph anomaly detection gets its generalization playbook","New review charts GAD's path from transfer to foundation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The survey's map is trustworthy only if the set of papers it reviews is representative of the field, but the paper gives no search strategy or inclusion criteria to guarantee that no significant generalized-GAD method is missing.","fun_headline_variants_meta":{"raw":{"variants":["Survey taxonomizes graph anomaly detection generalization","From transfer learning to foundation models: GAD survey","Two paradigms explain GAD generalization","Graph anomaly detection gets its generalization playbook","New review charts GAD's path from transfer to foundation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000416,"raw_usage":{"total_tokens":2143,"prompt_tokens":940,"completion_tokens":1203,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1134}},"tokens_in":556,"tokens_out":1203,"duration_ms":8797,"temperature":1.0,"reasoning_tokens":1134,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T16:14:24.043042+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Checking every generalized GAD paper published in the five years before the survey, one could ask whether each falls cleanly into exactly one of the four taxonomy leaves; finding a peer-reviewed method that fits none of them—or that the survey's own categories force into two places at once—would falsify the claim that the taxonomy is comprehensive and systematic.","supporting_citations":[{"cited_title":"Domain generalization: A survey.TPAMI, 45(4):4396–4415, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the notion of domain generalization that frames the paper's motivation for moving beyond same-distribution GAD."},{"cited_title":"A comprehensive survey on graph anomaly detection with deep learning.TKDE, 35(12):12012–12038, 2021","cited_arxiv_id":null,"evidence_quote":"Establishes the conventional GAD problem and its underlying assumptions, which the survey contrasts with generalized paradigms."},{"cited_title":"Deep graph anomaly detection: A survey and new perspectives.TKDE, 37(9):5106–5126, 2025","cited_arxiv_id":null,"evidence_quote":"Provides a recent broad survey of deep graph anomaly detection that the paper positions itself against as more focused on generalization."},{"cited_title":"Cross- domain graph anomaly detection.TNNLS, 33(6):2406–2415, 2021","cited_arxiv_id":null,"evidence_quote":"Pioneers cross-domain graph anomaly detection and serves as the anchor method for the transfer-learning branch."},{"cited_title":"Cross-domain graph anomaly detection via anomaly-aware contrastive alignment","cited_arxiv_id":null,"evidence_quote":"Exemplifies anomaly-aware one-class domain alignment, a key technique in the source-target representation alignment category."},{"cited_title":"Vicky Zhao, Yuan Yao, and Jia Li","cited_arxiv_id":null,"evidence_quote":"Unifies multi-level graph anomaly detection, representing the cross-granularity foundation model category."},{"cited_title":"Zero-shot generalist graph anomaly detection with unified neigh- borhood prompts.IJCAI, 2025","cited_arxiv_id":null,"evidence_quote":"Demonstrates zero-shot generalist graph anomaly detection via unified neighborhood prompts, anchoring the cross-scenario foundation model category."},{"cited_title":"Anomalygfm: Graph foundation model for zero/few-shot anomaly detection","cited_arxiv_id":null,"evidence_quote":"Provides a graph foundation model for zero/few-shot anomaly detection, supporting the cross-scenario and resource-efficient adaptation discussion."},{"cited_title":"ARC: A generalist graph anomaly detector with in-context learning","cited_arxiv_id":null,"evidence_quote":"Introduces a generalist graph anomaly detector with in-context learning, used as a central example of resource-efficient adaptation."},{"cited_title":"Bond: Benchmarking unsupervised outlier node detection on static attributed graphs","cited_arxiv_id":null,"evidence_quote":"Supplies a benchmark for unsupervised outlier node detection that the paper cites when calling for standardized evaluation protocols."}],"review_version":2}