{"id":"fc8ad561-9273-45d3-8075-631f9bbfddd3","arxiv_id":"2501.01124","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"LLM-for-graph methods are divided into Graph2text and Graph2token paradigms, with four conversion challenges and a model-selection guideline.","lead":"A survey proposes sorting research on large language models for graphs into two camps: turning graphs into human-readable text, or turning them into token-like vectors. It also identifies four difficulties in this conversion and gives practical tips for choosing models.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Graph2text/Graph2token dichotomy lacks an operational test: the paper's own definitions and assignments place hybrid methods and even a single method on both sides, so the central classification claim is not yet established.","rationale":"The paper's contribution is a two-way taxonomy plus four transformation challenges and a model-selection guideline. For the central claim to hold, methods must be classifiable into exactly one paradigm via the input format. That condition is insecure: Section 2.2's token definition conflates tokenizer atoms with graph-model embeddings, and the examples in Sections 4 and 5 do not follow a single criterion. GraphTranslator and Gpt4graph are concrete ambiguities in the assignments. This matters because the survey's subsequent organization, Figure 2, and practical guideline all inherit the taxonomy. It is not an outside-consensus dispute; it is an internal consistency issue. The reader's conditional verdict is appropriate: the finding is fixable, but only if the authors clarify the definitions and rework the classifications. The proposed test is a mechanical reclassification exercise, not an empirical benchmark, and would settle whether the concern lands.","tokens_in":37179,"tokens_out":5077,"duration_ms":50484,"concrete_test":"Build a decision tree from the definitions in Section 2.2 and Figure 2: (1) Is the LLM input a human-readable string (text) or a graph-derived embedding sequence (token)? (2) For hybrid inputs, define a tie-break rule. Then independently classify all methods listed in Sections 4 and 5. Report the number of methods that cannot be assigned uniquely, and specifically check GraphTranslator (5.1.1), Gpt4graph (4.1 vs 6.2.1), GraphPrompter (5.3.2), GraphQA (4.1 vs Table 1), and TokenGT (5.1.1). If two or more methods change category under alternative readings of 'token', the claimed exhaustive/mutually exclusive dichotomy fails and the taxonomy needs revision.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The taxonomy's central claim is that every LLM4graph method falls into exactly one of two paradigms, distinguished by whether the LLM input is text or tokens (Abstract, Section 1). This requires a stable criterion, but Section 2.2 defines 'tokens' in two incompatible ways: first as tokenizer subword units, then as numerical embeddings produced by auxiliary graph/neural models. Since every Graph2text input is also tokenized by the LLM, 'becomes tokens' cannot separate the paradigms; if 'token' means auxiliary embeddings, then methods that concatenate graph embeddings with natural-language instructions are hybrids that fit both sides. The assignments confirm the ambiguity: GraphTranslator is described in Section 5.1.1 as converting structured graph information into 'natural language instructions comprehensible to LLMs', yet it is categorized under Node2token (Graph2token); Gpt4graph is presented as a GDL/Graph2text method in Section 4.1 and as a soft-prompt example in Section 6.2.1 and Table 1. GraphPrompter (5.3.2) projects GNN embeddings into a vector space and combines them with text embeddings, so it spans both paradigms. Without a decision rule that assigns each method uniquely, the paper's central organizational claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript is a survey of methods that apply large language models (LLMs) to graph learning, termed LLM4graph. It argues that all such methods can be partitioned into two paradigms: Graph2text, which converts graphs into human-readable text, and Graph2token, which converts graphs into tokens or embeddings for LLM input. The paper identifies four transformation challenges (alignment, position, multi-level semantics, and context), uses them to organize the methods, provides practical guidance for model selection based on graph type, hardware, and optimization tricks, and outlines five future research directions.","tokens_in":37439,"tokens_out":9102,"duration_ms":76156,"significance":"The paper offers a broad and current collection of LLM4graph methods, a transformation-based perspective that goes beyond task- or architecture-centric surveys, and a practical resource table with hardware and code availability. If the taxonomy were precisely defined, this could become a useful reference for researchers entering the field. The main contribution, however, is the proposed Graph2text/Graph2token dichotomy, and that dichotomy currently lacks a clear operational boundary and contains internally inconsistent method assignments. The paper also compiles useful open problems, but the foundational classification claim needs substantial revision before the survey can serve as a reliable organizing framework.","major_comments":[{"comment":"The central dichotomy between Graph2text and Graph2token is not operationally defined. Section 2.2 defines 'tokens' in two incompatible ways: first as tokenizer subword units, then as numerical embeddings derived from auxiliary neural network processing. Since every Graph2text input is also tokenized by the LLM, the phrase 'becomes tokens' cannot separate the paradigms. The method assignments are also inconsistent: GraphTranslator is described in Section 5.1.1 as converting structured graph information into 'natural language instructions comprehensible to LLMs', yet it is categorized under Node2token; Gpt4graph is presented as a GDL/Graph2text method in Section 4.1 but appears as an example of soft prompts in Section 6.2.1 and Table 1; GraphPrompter in Section 5.3.2 feeds LLMs soft prompts that combine node embeddings with text embeddings, spanning both paradigms. The paper needs an explicit decision rule (for example, based on whether the LLM input is human-readable language tokens or continuous vectors inserted into the embedding space) and a reclassification of all methods according to that rule.","section":"Section 2.2; Section 5.1.1; Section 6.2.1; Table 1; Section 5.3.2"},{"comment":"The claimed use of the four challenges as subclassification criteria is inconsistent with the actual organization. Section 4 subclassifies Graph2text by graph type (general, AMR, KG) and then by challenges, whereas Section 5 subclassifies Graph2token by token granularity (node, pairwise, group, holistic) and then by challenges. Moreover, the challenge dimensions are unevenly covered: Section 5.2 (Pairwise nodes2token) has no Alignment subsection, and Section 5.3 (Group-aware nodes2token) has no Position subsection, while other categories contain different subsets of challenges. A consistent two-level taxonomy (for example, paradigm first, challenge second, then specific methods) is needed for the framework to be usable and for fair comparison across the two paradigms.","section":"Sections 4 and 5"}],"minor_comments":[{"comment":"The sentence 'This process addresses the problem of modality mismatch between graph data and text data [? ]' contains a placeholder citation '[? ]' that must be replaced with a proper reference.","section":"Section 5.1.1"},{"comment":"The sentence 'This process aims to align Graph2text properly to ensure that the structured knowledge is being inaccurately converted' is self-contradictory; 'inaccurately' should likely be 'accurately'.","section":"Section 4.3"},{"comment":"In the discussion of the context problem, 't is worth noting that texts are always regular sequences' is missing the initial 'I'.","section":"Section 3"},{"comment":"The phrase 'ensuring effective cross-modal intergration' contains a typo: 'intergration' should be 'integration'.","section":"Section 6.2.2"},{"comment":"The phrase 'perform consistenly' should be 'perform consistently'.","section":"Section 7.5"},{"comment":"The GraphPrompter row cites reference [127], but the correct reference for GraphPrompter is [125]; reference [127] is MoMu. Please correct the citation.","section":"Table 1"},{"comment":"The footnote explaining that 'text' refers both to a general modality and to the specific Graph2text representation is confusing; consider defining 'text' and 'token' in a single, unambiguous place, and avoid using 'textual representation' to cover both paradigms.","section":"Section 1 footnote"},{"comment":"The recommendation to use Pairwise nodes2token for heterogeneous graphs is not well supported by the methods cited there (N-GCN, T-PAIR, PLNLP, LPFormer, NAGphormer), which are mostly general link-prediction and node-classification methods; please justify or adjust the recommendation.","section":"Section 6.1.3"}],"recommendation":"major_revision","confidential_remarks":"The paper has the potential to be a useful survey for the community, but the central taxonomy needs substantial clarification and consistent application. The missing citation placeholder and several typographical errors suggest the manuscript is not yet polished. I would recommend the authors conduct a revision that sharpens the Graph2text/Graph2token boundary, reclassifies the methods that currently straddle the two paradigms, and applies the four challenges uniformly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, the Graph2text/Graph2token taxonomy is a genuinely useful way to organize the LLM-for-graph literature, and the four-challenge decomposition (alignment, position, multi-level semantics, context) is a sensible lens. Second, the paper never makes the two paradigms mutually exclusive, and several of its own method placements fall on both sides.\n\nWhat the paper does well: it covers a wider range of methods than most LLM4graph surveys, including the older AMR and KG-to-text literature that usually gets skipped. The practical guideline and Table 1, with hardware and open-source resources, are concretely useful. The five future directions are sensible. The survey is readable and the examples are well chosen.\n\nThe soft spots are real but fixable. Section 2.2 defines tokens first as subword units and then as graph-derived embeddings, and the paper doesn't reconcile the two. Because any Graph2text input is also tokenized, 'becomes tokens' cannot be the separator; if tokens mean auxiliary embeddings, then methods like GraphPrompter that concatenate text and graph embeddings are hybrids. The paper's own assignments show the ambiguity: GraphTranslator is described as converting graphs into natural-language instructions but placed under Node2token; Gpt4graph appears as a Graph2text GDL method in Section 4.1 and as a soft-prompt example in Section 6.2.1. There is also a contradictory sentence in Section 4.3 about 'ensuring that the structured knowledge is being inaccurately converted' where the opposite is clearly meant, and a missing citation placeholder in Section 5.1.1. No systematic survey methodology is documented, which makes the coverage hard to assess.\n\nNone of this sinks the survey. The taxonomy can be sharpened without changing its spirit, and for a practitioner wanting a map of the area, the paper is already useful. The authors should be asked to define a single criterion (e.g., whether the LLM input is primarily human-readable text or primarily model-derived embeddings), explicitly treat hybrid methods as a third category or explain a primary-assignment rule, and fix the internal contradictions.\n\nI'd send it to serious peer review. It's not a field-reshaper, but as a survey with a new organizational lens it deserves referee time. I wouldn't cite it until the boundary issues are resolved.","headline":"A useful survey framing undercut by a definitional blur: the Graph2text/Graph2token split is a good organizing idea but needs a clearer criterion before the taxonomy can be relied on.","tokens_in":37965,"tokens_out":2744,"would_cite":false,"duration_ms":26842,"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 every LLM-for-graph method is either Graph2text or Graph2token, defined by what the model receives.","keywords":["graph learning","large language models","Graph2text","Graph2token","graph-to-text transformation","survey","taxonomy","graph tokenization"],"falsifier":"The taxonomy would be settled by an annotation study: take a representative sample of published LLM4graph papers and ask independent annotators to classify each as Graph2text, Graph2token, or both/neither; if a substantial share lands in both/neither or annotators disagree, the dichotomy fails, and the paper's own placements of GraphTranslator and Gpt4graph are immediate candidates to test.","tokens_in":36947,"feed_emoji":"🕸️","tokens_out":6069,"duration_ms":55093,"temperature":0.7,"pith_summary":"This survey of LLM-based graph learning proposes that the entire method landscape is determined by one design choice: what the language model actually receives. The paper's central claim is that every existing method is either Graph2text, which serializes a graph into human-readable strings, or Graph2token, which converts nodes, edges, or subgraphs into token or embedding inputs. It argues that both paradigms confront the same four transformation problems—alignment, position, multi-level semantics, and context—and organizing the literature this way explains why and how LLMs can reason over irregular graph structures. The paper also draws practical conclusions: which paradigm fits which graph type, when prompting versus fine-tuning is cheaper or safer, and which open problems remain.","feed_headline":"LLM graph learning splits into just two paradigms","feed_subtitle":"A new survey groups all methods by how the graph is transformed before the model sees it: text or tokens.","key_machinery":"The central object is the transformation step from graph to LLM input, and the paper's machinery is a two-armed taxonomy: Graph2text versus Graph2token, with the latter subdivided by token granularity (node, pairwise, group-aware, holistic). Four named challenges—alignment, position, multi-level semantics, and context—serve as the analytic grid that assigns each subclass its problem. The taxonomy carries the argument by converting an unstructured method zoo into a two-by-four design space that the paper then uses to give model-selection and prompting-versus-fine-tuning guidance.","core_discovery":"The paper's central claim is taxonomic: when a graph is fed to an LLM, the format of that input is the fundamental dividing line in the field the authors call LLM4graph. Graph2text covers methods that translate graph structure into natural language or graph-description languages such as GraphML and query languages like Cypher, preserving interpretability and using the LLM's text-processing strength. Graph2token covers methods that encode graph components as tokens—single nodes, node pairs, node groups, or the whole graph—often via a graph encoder or projection layer, so that structural and semantic information is fused before the LLM sees it. The paper pairs this bipartition with four challenges of graph-to-text transformation (alignment, position, multi-level semantics, context) and uses them as a problem-oriented grid to subclassify and compare methods. If the taxonomy is right, the scattered toolbox of LLM-graph techniques reduces to a small set of design decisions, and the four challenges become a checklist for evaluating any new method.","pith_inferences":["Beyond the paper, the Graph2text/Graph2token distinction is best read as a spectrum: many deployed systems already mix natural-language prompts with projected graph embeddings, so the exhaustive dichotomy is a useful approximation rather than a sharp law.","Beyond the paper, the four challenges could be turned into a falsifiable evaluation protocol: report which challenges each method addresses and measure the performance gain when each is explicitly handled.","Beyond the paper, the suggestion to serialize graphs with database query languages (Cypher, Gremlin) is directly testable by comparing LLM accuracy on identical graph reasoning tasks across serialization formats.","Beyond the paper, the treatment of non-edges as negative semantics suggests a concrete experiment: check whether Graph2token models that explicitly encode disconnected node pairs beat those that encode only observed edges on link prediction benchmarks."],"forward_implications":["Any future LLM4graph method can be classified by a single question: does the LLM receive readable text or token/embedding input?","The four challenges give a concrete checklist—alignment, position, multi-level semantics, context—for designing graph serializations and for spotting why a given method underperforms.","The practical guideline says Graph2text suits textual graphs, AMR graphs, and knowledge graphs where interpretability matters, while Graph2token suits attributed, heterogeneous, community, and molecular graphs where structure must be fused into representations.","Prompt-based tricks are preferred when data is scarce or when privacy rules out sending sensitive graph data to external APIs; fine-tuning is preferred when the task demands granular control and customization.","The paper's five open problems (general instructions, theoretical invariance, fairness, efficiency and scalability, dynamic graphs) give a roadmap of where the taxonomy stops being prescriptive."],"supporting_citations":[{"why":"Provides the natural-language serialization of graph nodes and edges that defines the Graph2text/NL branch.","marker":"[53]"},{"why":"Supplies algorithmic natural-language prompting for graph reasoning, a canonical Graph2text baseline the survey builds its taxonomy around.","marker":"[54]"},{"why":"Demonstrates graph-description-language input and is cited in both the Graph2text GDL branch and the prompting discussion.","marker":"[56]"},{"why":"Instantiates the GDL-program conversion that grounds the Graph2text GDL subcategory.","marker":"[58]"},{"why":"Uses random-walk-based graph-to-text fine-tuning, a load-bearing example for Graph2text alignment and position on knowledge graphs.","marker":"[76]"},{"why":"Grounds Node2token alignment with a node dictionary that unifies GNN and text token spaces.","marker":"[99]"},{"why":"Treats nodes and edges as identical tokens, the purest example of the Graph2token paradigm.","marker":"[100]"},{"why":"Supplies a text-to-graph contrastive alignment mechanism used to define Graph2token alignment.","marker":"[32]"},{"why":"Exemplifies the Node2token alignment route via a translator module that maps graph structure into LLM-comprehensible instructions.","marker":"[101]"}],"fun_headline_variants":["Graph2text vs Graph2token: LLM graph divide","Survey: Only two ways to feed a graph to an LLM","New taxonomy splits LLM graph methods into text or token","LLM4graph: all methods fall into text or token input","Two paradigms for LLMs learning from graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The dichotomy between Graph2text and Graph2token is exhaustive and mutually exclusive: every LLM4graph method can be assigned unambiguously to exactly one paradigm based on the format of the LLM input.","fun_headline_variants_meta":{"raw":{"variants":["Graph2text vs Graph2token: LLM graph divide","Survey: Only two ways to feed a graph to an LLM","New taxonomy splits LLM graph methods into text or token","LLM4graph: all methods fall into text or token input","Two paradigms for LLMs learning from graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000237,"raw_usage":{"total_tokens":1510,"prompt_tokens":949,"completion_tokens":561,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":565,"tokens_out":561,"duration_ms":5591,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:34:05.817556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The taxonomy would be settled by an annotation study: take a representative sample of published LLM4graph papers and ask independent annotators to classify each as Graph2text, Graph2token, or both/neither; if a substantial share lands in both/neither or annotators disagree, the dichotomy fails, and the paper's own placements of GraphTranslator and Gpt4graph are immediate candidates to test.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Grounds Node2token alignment with a node dictionary that unifies GNN and text token spaces."}],"review_version":1}