{"id":"a516cb51-9fbd-4605-96f5-c2ad43ca67a4","arxiv_id":"2506.08326","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A survey of graph prompting methods that categorizes them by the stage at which prompts are applied: data, representation, or task.","lead":"This paper reviews recent techniques for adapting pre-trained graph models to new tasks using trainable prompts, and organizes them into a taxonomy based on data, representation, and task levels. It serves as a reference for researchers entering graph prompt learning.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The novelty claim for the taxonomy is never substantiated: Section 1 asserts a novel data/representation/task split 'compared to previous survey papers [61,87]' but no comparison is provided, so the paper's central contribution may be relabeling an existing scheme.","rationale":"We read the survey in good faith: it covers a wide range of graph prompting methods, organizes them under a clear Figure 2, and includes pre-training background, applications, and future directions. That breadth supports the 'systematic review' part of the central claim. The load-bearing weakness is the explicit novelty claim. Because the paper is a survey, its main intellectual contribution is the taxonomy; the claim that the taxonomy is novel is both asserted and checkable. It is not enough to say 'compared to [61,87]' without showing the comparison. The fix is straightforward: add a comparison table or paragraph. On the reader's GPF-plus concern, I see a real ambiguity in Section 4.1.1: the phrase 'independent prompt vector to each node' conflicts with Eq. (6), where p_i is generated from shared basis vectors and node-dependent softmax weights. If the original method indeed uses per-node free parameters, Eq. (6) is wrong; if it uses basis-vector generation, the prose is misleading. I did not make this the primary concern because the conflict can be resolved by reading the original paper and does not directly threaten the taxonomy's novelty. The CONDITIONAL verdict is appropriate: the survey is useful and fixable, but the headline contribution should be verified against the cited surveys before the paper can be relied on as a definitive reference.","tokens_in":19884,"tokens_out":9308,"duration_ms":108409,"concrete_test":"Obtain the full texts of [61] and [87], extract their top-level taxonomy definitions, and tabulate where each survey places the methods that appear in this paper's Table 1. If either prior survey already partitions graph prompting into data-level, representation-level, and task-level categories (or categories that differ only in name), then the Section 1 novelty bullet should be withdrawn or explicitly qualified. If both prior surveys use orthogonal axes, add a one-paragraph comparison showing the distinction, and the novelty claim can stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated contribution (Section 1, bullet 1) is a 'novel taxonomy' of graph prompting methods from the perspectives of graph data, node representations, and downstream tasks, and the text says this is proposed 'Compared to previous survey papers [61,87].' This novelty claim is load-bearing because the survey's value is organizational: if an equivalent tripartite split already exists in [61] (arXiv:2408.14520) or [87] (arXiv:2311.16534), then the main contribution is not achieved. The survey never defines the earlier taxonomies, never gives a comparison table, and never states a distinguishing criterion. The absence is especially salient because both cited surveys have overlapping titles and likely overlapping coverage; without a side-by-side mapping the reader cannot tell whether 'data-level / representation-level / task-level' is a new axis or a relabeling. This is a checkable factual claim about two cited documents, not a disagreement with consensus. Section 6.1-6.5 discuss open challenges but do not revisit the novelty assertion. If the split is not new, the paper's primary contribution should be restated as 'an up-to-date survey organized by a convenient three-way split,' which is a weaker but still useful claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This survey reviews graph prompting, the approach of adapting pre-trained graph learning models to downstream tasks by learning trainable prompts while keeping the model frozen. It first summarizes graph pre-training foundations (generative, contrastive, and multi-task methods), then proposes a taxonomy of graph prompting techniques into data-level, representation-level, and task-level prompting. It also surveys applications in recommendation, knowledge engineering, biology, and medicine, and closes with open challenges and future directions. The paper's stated contribution is organizational: a novel taxonomy and an up-to-date compilation of recent methods.","tokens_in":20173,"tokens_out":4056,"duration_ms":48111,"significance":"If accurate, the survey would be a useful and timely reference for researchers entering graph prompting, and its three-level taxonomy (data, representation, task) is a plausible organizing scheme. The paper is broad and current, covering both foundational pre-training methods and very recent prompting techniques, and it includes a helpful summary table. Because the contribution is organizational, the accuracy of method descriptions and the distinctiveness of the taxonomy are load-bearing; the current inconsistencies and the unsubstantiated novelty claim diminish but do not destroy the survey's value. The paper contains no derivations, so there is no circularity concern; self-citations are used as ordinary references to relevant prior work.","major_comments":[{"comment":"The novelty claim for the taxonomy is not substantiated. The text says the data/representation/task split is proposed 'Compared to previous survey papers [61,87]', but the paper never describes the taxonomies in either survey, never provides a comparison table, and never states a distinguishing criterion. Since this is the paper's primary stated contribution, the authors should either provide a side-by-side comparison showing what the new taxonomy adds, or restate the contribution as an up-to-date survey organized by a convenient three-way split. This is a concrete, checkable claim about two cited documents, not a matter of taste.","section":"Section 1, first contribution bullet"},{"comment":"The description of GPF-plus is internally inconsistent. The text first says GPF-plus 'assigns an independent prompt vector to each node', then says it uses attentive aggregation of basis vectors, and Eq. (6) defines each prompt as a weighted combination of shared basis vectors. An independent per-node vector and a shared basis set with per-node weights are different mechanisms. Please clarify which mechanism is being described and reconcile the surrounding prose with Eq. (6); as written, a reader cannot tell what GPF-plus actually does.","section":"Section 4.1.1, Eq. (6)"},{"comment":"Several methods appear in more than one category without explanation. GraphPrompt appears under both Output representation prompting and Similarity-based prompting; GraphPrompt+, HetGPT, and ProNoG similarly span representation-level and task-level rows. Section 4 introduces the three categories as a division of techniques, which implies they are mutually exclusive at the method level. If a method legitimately operates at multiple levels, the paper should say so explicitly and mark primary versus secondary categorizations; otherwise the taxonomy's consistency is undermined.","section":"Table 1"}],"minor_comments":[{"comment":"The reference list contains duplicate entries for the same works: [9] and [10] are both the GPF paper, [19] and [20] are both GraphSAGE, [118] and [119] are both GraphCL, [124] and [125] are the same non-homophilic graph paper in preprint and published form, and [138] and [139] are both the CoOp paper. These should be merged.","section":"References"},{"comment":"The abbreviation 'DT Universality' is not defined anywhere; please spell it out in the caption or in the text.","section":"Table 1"},{"comment":"The sentence 'Therefore, no additional prompts are introduced in similarity-based prompting' conflicts with the immediately following sentence stating that some studies design class prototypes as learnable vectors. Please rephrase to distinguish prototype-averaging methods from learnable-prototype methods.","section":"Section 4.3.2"},{"comment":"The phrase 'an attribute graph' should be 'an attributed graph'.","section":"Section 2.1"},{"comment":"The taxonomy diagram is dense and would benefit from a legend or from matching color/shading to the three prompting levels, since the table and text use the same categories.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a survey venue and the topic is timely. The main issue is that the headline novelty claim is currently unverifiable because the taxonomies of [61] and [87] are never compared; this is fixable with a comparison table. The GPF-plus inconsistency should also be corrected before the survey can be relied on as a reference. I would not recommend rejection, but the revision must address the load-bearing points above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent survey of graph prompting, and the data/representation/task split is a workable way to organize the field. It covers the pretraining-prompting pipeline clearly, includes work up to 2025, and the applications section is broad. If you are new to graph prompting, this gives you a solid map of the area. The main soft spot is the novelty claim. Section 1 says the taxonomy is proposed compared to previous survey papers [61,87] but never tells the reader what those surveys' taxonomies are, never gives a comparison table, and never states a distinguishing criterion. That is a checkable factual claim about two cited documents, and without the comparison the claim does not hold up. The honest contribution is an up-to-date survey organized by a convenient three-way split, which is weaker but still useful. There are also mechanical errors that reduce the paper's value as a reference. Duplicate reference entries ([9]/[10], [19]/[20], [138]/[139]) suggest sloppy final proofing. The GPF-plus description in Section 4.1.1 says it further assigns an independent prompt vector to each node and then immediately describes the basis-vector aggregation that makes it parameter-efficient; that is internally inconsistent on its face. And Table 1 lists GraphPrompt in both the output-representation and similarity-based rows without noting that the method deliberately does both. None of these are load-bearing enough to sink the survey, but they are exactly the kind of thing that matters when a survey's job is to be a trusted pointer to the literature. I would send this to peer review. A referee should ask for the taxonomy comparison with [61,87] and a fix of the GPF-plus inconsistency and reference duplicates. If the authors make those changes, this is a solid survey that people in the graph learning community will cite.","headline":"A useful, up-to-date survey of graph prompting whose headline novelty claim is unsubstantiated because it never compares its taxonomy with the two earlier surveys it cites.","tokens_in":586,"tokens_out":810,"would_cite":true,"duration_ms":34308,"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 that the scattered literature on graph prompting can be organized into a single three-part taxonomy—where the prompt acts on the input graph, on node representations, or on the downstream task—and uses that taxonomy to…","keywords":["graph prompting","graph pre-training","graph neural networks","prompt tuning","taxonomy","few-shot learning","self-supervised learning"],"falsifier":"A concrete refutation would be a graph prompting method whose prompt acts across all three levels in a way that cannot be represented by the taxonomy, or a documented demonstration that the dividing lines here match the taxonomy of the two earlier surveys the paper itself cites; either would show the claimed novelty or exhaustiveness does not hold.","tokens_in":19733,"feed_emoji":"🧭","tokens_out":5041,"duration_ms":57457,"temperature":0.7,"pith_summary":"This survey attempts to organize the fast-growing literature on graph prompting, the practice of adapting a pre-trained graph neural network to new tasks by learning small trainable prompt vectors while keeping the model itself frozen. It proposes that all existing graph prompting methods can be sorted into three categories according to where the prompt acts: on the input graph data, on node representations inside the model, or on the downstream task itself. If this taxonomy holds, it gives researchers a shared vocabulary for comparing methods that currently look unrelated, and a way to see which combinations of prompt levels have not yet been tried. The survey also catalogues the graph pre-training methods that prompts build on, surveys applications in recommendation, knowledge graphs, and biomedicine, and identifies missing theory, benchmarks, robustness, and LLM integration as open problems.","feed_headline":"Graph prompting gets a three-level map: data, model, task","feed_subtitle":"A new review sorts every graph prompting method by where it acts, and spells out the gaps that remain.","key_machinery":"The taxonomy itself is the carrying mechanism: the distinction between data-level prompting (transforming the input graph into a prompted graph), representation-level prompting (transforming the hidden or output representation matrix at a chosen layer), and task-level prompting (transforming the downstream task into a different form). Each category is tied to a concrete mathematical operation on a different stage of the pipeline, which is what makes the organizing scheme more than a list. The survey further splits data-level prompting into feature-based and insertion-based strategies, representation-level prompting into output and hidden variants, and task-level prompting into link-prediction-based and similarity-based variants.","core_discovery":"The central organizational claim is that the entire space of graph prompting techniques can be captured by three categories defined by where the learnable prompt is inserted. Data-level prompting modifies the input graph, either by adding a prompt vector to node features or by inserting prompt nodes into the graph. Representation-level prompting multiplies output or hidden node representations by prompt vectors, using either a shared prompt or node-wise and layer-wise prompts. Task-level prompting reformulates the downstream problem, typically converting node classification into link prediction or into similarity matching against class prototypes. The paper presents this three-way split as a novel taxonomy that improves on earlier reviews, and uses it to structure a comprehensive summary of recent methods, their applications, and open challenges.","pith_inferences":["If the taxonomy is right, the prompt level and the pre-training task form an implicit design matrix; many cells of that matrix, such as hidden-representation prompts applied to a multi-task pre-trained model, remain untested, and filling them would be a cheap empirical programme.","The data-level prompt update rule resembles an adversarial perturbation under a norm constraint, which suggests that the adversarial robustness questions the survey raises could be studied with existing tools from adversarial example analysis, an extension the paper does not make.","A testable prediction follows from the taxonomy: methods in the same category should show similar failure modes and transfer behaviour across datasets, whereas methods in different categories should fail differently, and a meta-analysis of published results could check this without new experiments."],"forward_implications":["A researcher facing a new graph task can locate existing prompts by asking where the prompt acts in the pipeline, making method selection a mapping task rather than a literature search.","Because data-level prompting requires a full forward and backward pass while task-level prompting does not, the taxonomy exposes a cost and compatibility trade-off that future prompt designs can optimize.","The taxonomy suggests that prompts at different levels are complementary, so combinations such as feature-based prompts with similarity-based task reformulation are a natural next direction.","Task-level prompting is currently limited to classification; the survey's framing implies that regression and other task types still lack a principled prompting formulation.","The review redirects the field's open challenges to benchmark standardization, theoretical foundations, universal compatibility, robustness, and LLM incorporation."],"supporting_citations":[{"why":"A prior survey on graph prompt learning whose taxonomy this paper claims to improve upon.","marker":"[61]"},{"why":"Another prior comprehensive survey on graph prompt learning that this paper contrasts with.","marker":"[87]"},{"why":"Introduces GPF and GPF-plus, foundational feature-based data-level prompting methods.","marker":"[9]"},{"why":"Introduces GraphPrompt, a key output representation prompting method also used in similarity-based task prompts.","marker":"[59]"},{"why":"Introduces GPPT, the main link-prediction-based task-level prompting method.","marker":"[84]"},{"why":"Introduces All-in-one, a representative insertion-based data-level prompting method with multi-task prompting.","marker":"[86]"},{"why":"ProG is the only benchmark cited that standardizes evaluation of graph prompting methods.","marker":"[150]"},{"why":"GraphMAE exemplifies the generative pre-training approach that serves as the foundation stage for graph prompting.","marker":"[24]"},{"why":"Deep Graph Infomax exemplifies contrastive graph pre-training, another foundation for graph prompting.","marker":"[95]"}],"fun_headline_variants":["Graph prompting taxonomy: data, representation, task levels","Three-level map for graph prompting: where to insert prompts","Survey sorts graph prompting by where prompts act","Graph prompting reviewed: three categories, gaps remain","Prompting graph models: a structured review of advances"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the proposed three-category taxonomy is genuinely new and faithfully describes every method it places, because the paper's main contribution is organizational and would weaken if earlier surveys already draw the same dividing lines or if some methods are misdescribed.","fun_headline_variants_meta":{"raw":{"variants":["Graph prompting taxonomy: data, representation, task levels","Three-level map for graph prompting: where to insert prompts","Survey sorts graph prompting by where prompts act","Graph prompting reviewed: three categories, gaps remain","Prompting graph models: a structured review of advances"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000273,"raw_usage":{"total_tokens":1584,"prompt_tokens":842,"completion_tokens":742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":668}},"tokens_in":458,"tokens_out":742,"duration_ms":7743,"temperature":1.0,"reasoning_tokens":668,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:12:38.831046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete refutation would be a graph prompting method whose prompt acts across all three levels in a way that cannot be represented by the taxonomy, or a documented demonstration that the dividing lines here match the taxonomy of the two earlier surveys the paper itself cites; either would show the claimed novelty or exhaustiveness does not hold.","supporting_citations":[{"cited_title":"Towards Graph Prompt Learning: A Survey and Beyond","cited_arxiv_id":"2408.14520","evidence_quote":"A prior survey on graph prompt learning whose taxonomy this paper claims to improve upon."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ProG is the only benchmark cited that standardizes evaluation of graph prompting methods."}],"review_version":1}