{"id":"af4d133c-e8c1-461b-af83-677d8a7808f7","arxiv_id":"2506.11012","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A task-oriented survey of knowledge graph reasoning, covering six task categories, benchmark datasets, downstream applications, and future challenges.","lead":"This survey organizes hundreds of knowledge graph reasoning methods into six task types and reviews their applications and open challenges. It serves as a structured map for researchers and practitioners who want to understand how machines infer new facts from structured knowledge.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The six-task taxonomy mixes criteria (hop count, dynamics, modality, data scarcity, inductive setting) and is not a partition; the survey's comprehensiveness claim lacks a documented selection or coverage protocol.","rationale":"The paper is a genuinely broad survey with a useful map of KGR models, datasets, applications, and open problems, and the GitHub repository adds a degree of transparency. My concern is not with individual model descriptions; rather, it is with the central framing. The abstract's 'more comprehensive perspective' requires that the six task families be a principled partition of the literature. The text shows they are not: TITer and FITCARL are each discussed under two different primary tasks, and the categories mix structural, temporal, modal, scarcity, and inductive criteria. This does not destroy the survey's value—overlapping scenario axes can be a legitimate organizing device—but it undercuts the implied exhaustiveness. The absence of a documented selection protocol (databases, search dates, inclusion criteria) further prevents readers from checking whether any significant body of work was omitted. These are addressable framing and methodology issues, consistent with a conditional acceptance. I therefore leave the reader's CONDITIONAL verdict unchanged.","tokens_in":46507,"tokens_out":4215,"duration_ms":45404,"concrete_test":"Run an inter-annotator classification exercise on 100 papers from 2022–2025 (drawn from the survey's own reference list plus a fresh keyword query for 'knowledge graph reasoning' and 'knowledge graph completion'). Have two annotators independently assign each paper to one of the six primary categories, with explicit options 'multiple' and 'other'. Compute Cohen's kappa and the fraction of 'other'/'multiple' labels. If kappa is below 0.6 or more than 15% of papers fall outside a single category, the six-way taxonomy is not a natural exhaustive partition, and the survey must state a priority rule or adopt a multi-label matrix before its comprehensiveness claim can be accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that this survey is the first to systematically cover all KGR tasks along with downstream applications and challenging paradigms—depends on the six categories in Section I-A and Fig. 1 being a natural, comprehensive way to organize the literature. That condition is not met as stated. The categories are defined on different axes: single-step vs multi-step is a reasoning-complexity axis; dynamic is a KG-type/temporal axis; multi-modal is an input-modality axis; few-shot is a data-availability axis; inductive is a generalization-setting axis. These axes intersect, so a single work can be assigned to several 'primary' tasks. The text itself double-lists models: TITer [243] appears in V-B4 (multi-step TKGR, under dynamic KGR) and again in VIII-C3 (temporal-enhanced IKGR); FITCARL [310] appears in VII-E2 and VIII-C3. Without a priority rule or a multi-label/matrix organization, the taxonomy cannot be exhaustive or mutually exclusive. Moreover, the paper gives no systematic literature-selection protocol, so the claimed gap over [23]–[33], summarized only as binary checks in Table I, is not independently verifiable. A survey organized around intersecting scenarios can still be useful, but its 'more comprehensive' contribution is weaker than claimed until these boundary and coverage questions are resolved.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents a task-oriented survey of knowledge graph reasoning (KGR). It organizes the literature into six primary task families—static single-step, static multi-step, dynamic, multi-modal, few-shot, and inductive KGR—and devotes one section to each, reviewing representative methods with per-section taxonomy figures. The survey also compiles benchmark datasets, discusses downstream applications in both horizontal domains (question answering, recommendation, visual reasoning) and vertical domains (healthcare, business, cybersecurity, education, transport, environment), and closes with challenges and opportunities including sparse KGR, uncertain KGR, KG error detection, trustworthy KGR, and LLM-enhanced KGR. The central claim is that, in contrast to prior surveys, this work provides a more comprehensive perspective by jointly covering all six KGR task families along with downstream applications and challenging reasoning paradigms.","tokens_in":46716,"tokens_out":5509,"duration_ms":61898,"significance":"If the organizing taxonomy is accepted, the survey has clear value as a reference: it collects a large body of recent work, including many 2023–2024 LLM-based methods, provides structured per-category reviews, summarizes widely used datasets in tables, and gives an application-oriented view that most prior KGR surveys lack. The companion GitHub repository and the per-section taxonomy figures are useful navigation aids. The principal risk is not the accuracy of individual method descriptions but the defensibility of the 'more comprehensive' claim: the six-task taxonomy is built on intersecting criteria, and the comparison with prior surveys is presented without a verifiable coverage protocol.","major_comments":[{"comment":"The six task classes are defined along different conceptual axes: single-step versus multi-step is a reasoning-complexity axis, dynamic is a temporal KG axis, multi-modal is an input-modality axis, few-shot is a data-availability axis, and inductive is a generalization-setting axis. These axes intersect, so the categories do not form a partition of the literature. The text itself demonstrates the overlap by discussing TITer in both Section V-B4 and Section VIII-C3, and FITCARL in both Section VII-E2 and Section VIII-C3. Since the abstract claims to categorize approaches 'based on primary reasoning tasks,' the paper should either state an explicit assignment or priority rule that determines the primary task of each work, or reframe the organization as a multi-label/matrix structure with a discussion of how intersections are handled. Without this, the claim that the taxonomy systematically organizes the full KGR literature is not established.","section":"Sections I-A, Fig. 1, and Sections V-B4/VII-E2/VIII-C3"},{"comment":"The comparison with prior surveys [23]–[33] is presented as binary checkmarks, but the paper provides no selection protocol, search strategy, inclusion/exclusion criteria, or operational definitions for the rows such as 'Downstream Tasks' and 'LLM.' A reader cannot verify the central claim that previous surveys lack these elements or that this survey uniquely covers all six task families. Please add a methodology paragraph describing how the surveyed papers were collected and how Table I was compiled, ideally with representative citations or section numbers from the compared surveys for each checkmark. This is load-bearing because the paper's distinctive contribution is precisely the claimed gap over prior reviews.","section":"Section I-B, Table I"},{"comment":"The paper asserts that KGR tasks 'can be categorized into six types' but gives no argument that these six are exhaustive. The text itself discusses zero-shot KGR as a distinct setting (e.g., ZSGAN, OntoZSL, DOZSL, DMoG in Section VII-C) and treats sparse KGR and uncertain KGR as challenges in Section X rather than as primary tasks. If zero-shot KGR is a separate scenario from few-shot KGR, or if sparse and uncertain KGs define distinct reasoning settings, the six-type taxonomy is incomplete. Please justify why these are excluded from the primary-task list, or explicitly state that the six categories are organizing scenarios rather than a formal exhaustive partition.","section":"Section II-B and Section VII-C"}],"minor_comments":[{"comment":"The rule-instantiation example is internally inconsistent: the observed triples are (David, BornIn, San Francisco) and (San Francisco, LocatedIn, U.S.A.), but the rule is instantiated as (David, BornIn, California) and (California, LocatedIn, U.S.A.). Section II-B(2) also gives the path David → San Francisco → California, which would suggest Nationality(David, California), conflicting with Nationality(David, U.S.A.). Please make the example triples and the inferred conclusion consistent throughout.","section":"Section III, introductory paragraph"},{"comment":"The notation for the support and query sets is garbled: 'Sr ={(hi, r, ti )|(hi, r, ti )∈ Tr}(i = 1) K' and the displayed expression for Qr do not parse. Please rewrite these definitions with proper subscript/superscript ranges.","section":"Section VII, few-shot setup"},{"comment":"The test-set relation count for DBPedia50k is listed as '-'; either provide that statistic or explain why it is omitted, since all other rows in the table report both entity and relation counts.","section":"Section IX-A, Table VII"},{"comment":"There are numerous typographical errors that should be corrected in a revision, including 'techiniques' (Section IX-C4), 'drivend' (Section I-A), 'cannnot' (Section III-A1), 'timestemps' (Sections V-B1 and V-B2), and 'entites' (Section III-A1).","section":"Throughout"},{"comment":"KoPA is cited as both [84] and [275]; please unify duplicate references to the same work.","section":"References [84] and [275]"},{"comment":"The text refers to 'TLIP' when describing temporal rule learning models, but the cited model is TILP [236]; please correct the name for consistency.","section":"Section X-D"}],"recommendation":"major_revision","confidential_remarks":"The survey is broad and likely useful to the KGR community, but the core contribution—the 'more comprehensive' taxonomy and the comparison with prior surveys—needs to be made defensible. I recommend major revision rather than rejection because the material is valuable and the concerns are addressable through reframing the taxonomy as an intersecting multi-axis organization, adding a coverage methodology, and fixing the expository inconsistencies."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a competent survey that gives newcomers a genuinely useful map of knowledge graph reasoning by task type, with benchmarks, applications, and challenges. The central claim—being the first to systematically cover all six task families plus downstream applications—is overstated, because the authors' own Table I shows prior surveys covered each task individually, and the six-way taxonomy is not a clean partition. That said, the integrated presentation has real value.\n\nWhat's actually new: the organization around six reasoning scenarios (static single/multi-step, dynamic, multi-modal, few-shot, inductive) plus dedicated sections on applications and open challenges is broader than prior surveys. The benchmark tables and taxonomy figures are well designed, and the writing is generally clear. Coverage is current through 2024, including LLM-based methods. The GitHub repo is a nice companion. Credit is due for the breadth and effort.\n\nThe soft spots are real but manageable. The stress-test concern is correct: the categories mix criteria. Single-step vs multi-step is reasoning complexity; dynamic is a temporal KG type; multi-modal is input modality; few-shot is data availability; inductive is a generalization setting. These axes overlap. A single work can land in multiple primary tasks, and the text itself double-lists TITer (V-B4 and VIII-C3) and FITCARL (VII-E2 and VIII-C3). Without a priority rule or multi-label matrix, the taxonomy is an organizing device, not a partition. The lack of a documented literature selection protocol also makes the coverage claim hard to verify independently; binary checks in Table I don't prove comprehensiveness.\n\nThere are minor quality issues: the example in Section III has a concrete error—the rule instantiation writes BornIn(David, California) where it should be BornIn(David, San Francisco)—plus scattered typos (e.g., 'timestemp'). These are cosmetic but worth fixing.\n\nWho it's for: newcomers wanting a map of the field, researchers looking for a single entry point to tasks, benchmarks, and applications. It doesn't resolve a scientific question, but it is an honest synthesis. My recommendation: send it to peer review. A serious referee can check the taxonomy, clean up the double-listings, and push the authors to either justify the six axes or soften the comprehensiveness claim. With reasonable revision, it would be a solid survey contribution.","headline":"A useful, broad task-oriented KGR survey whose main soft spot is an unproven taxonomy; worth a careful referee, not a desk reject.","tokens_in":47260,"tokens_out":1778,"would_cite":true,"duration_ms":18894,"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 full landscape of knowledge graph reasoning can be organized into six task families, and that earlier surveys missed the integration of downstream applications and open challenges.","keywords":["knowledge graph reasoning","task-oriented taxonomy","static single-step reasoning","dynamic reasoning","multi-modal reasoning","few-shot reasoning","inductive reasoning","large language models"],"falsifier":"Count the KGR papers published in the last three years, assign each to exactly one of the six task categories, and require that every assignment be unambiguous; if a sizable residue falls outside all six categories (e.g., hyper-relational or commonsense-only KGs, or zero-shot without a few-shot support set), the exhaustiveness claim is refuted.","tokens_in":46266,"feed_emoji":"🧠","tokens_out":5707,"duration_ms":53068,"temperature":0.7,"pith_summary":"This survey tries to give a single organizing map of knowledge graph reasoning (KGR), the task of inferring new facts from the facts already stored in a knowledge graph. It sorts the field into six task families—static single-step, static multi-step, dynamic, multi-modal, few-shot, and inductive reasoning—and reviews the main model strategies in each. Its claim to novelty is scope: earlier reviews covered one or a few of these families, while this one also ties each family to benchmark datasets, downstream applications (question answering, recommendation, visual reasoning, healthcare, business, cybersecurity), and open challenges such as sparse and uncertain graphs. If the taxonomy holds, a researcher or practitioner can use the survey as a map from task to method to application.","feed_headline":"Six task families organize knowledge graph reasoning research","feed_subtitle":"The survey unifies static, dynamic, multi-modal, few-shot, and inductive reasoning with downstream uses and open problems.","key_machinery":"The load-bearing object is the six-task taxonomy, defined by KG type and reasoning scenario, together with the cross-task map in the survey's overview figure, where numbered circles mark shared ideas (e.g., logic-rule learning in static single-step, temporal, and inductive KGR; meta-learning in few-shot and temporal few-shot). The taxonomy does the work of delimiting the field: each task section groups models by mechanism and ends with an advantage/limitation comparison, and the same structure is reused to organize benchmarks, applications, and open challenges.","core_discovery":"On its own terms, the paper's central claim is that KGR research is best understood through a task-oriented taxonomy with six primary tasks, and that no previous survey integrated all six with downstream applications and harder reasoning paradigms. The paper argues that static single-step KGR (predicting one missing element of a triple) and static multi-step KGR (following paths) are the two basic static scenarios; dynamic KGR adds updates and timestamps; multi-modal KGR adds text, images, and video; few-shot KGR handles long-tailed relations with few support triples; and inductive KGR handles unseen entities and relations at test time. Within each family, it categorizes the model landscape (e.g., embedding-based vs. rule-based for single-step, random walk vs. reinforcement learning vs. LLM for multi-step, time embedding vs. evolution learning for temporal), and it identifies cross-task shared ideas, such as rule learning appearing in static, temporal, and inductive settings. The survey also positions LLM-based reasoning as an emerging cross-cutting technique rather than a separate task.","pith_inferences":["Beyond the paper, the taxonomy invites a direct test: map recent KGR papers onto the six categories and see whether hybrid papers (e.g., temporal few-shot inductive reasoning) force a cross-product view rather than six clean boxes.","If LLM-based agents become the dominant interface to knowledge graphs, the static/dynamic/multi-modal distinction may shift from being about model family to being about data modality and time axis, which would change how the field is partitioned.","The survey's emphasis on downstream applications implies a benchmark design principle: evaluate KGR models not only by MRR and Hits@N, but by their effect on end tasks such as question-answering accuracy or recommendation explainability.","The paper's separation of few-shot and inductive reasoning suggests a testable boundary: few-shot handles limited observed triples for known relations, inductive handles unseen entities; work doing both at once would straddle the line."],"forward_implications":["A newcomer can use the six-task taxonomy to locate any KGR method and its closest alternatives.","Methods developed for one task can be transplanted to another where the numbered circles in the overview show shared ideas, such as rule learning appearing in static, temporal, and inductive settings.","Benchmarks and applications in question answering, recommendation, healthcare, finance, and cybersecurity give practitioners a direct path from model choice to deployment.","LLM-based KGR is treated as a cross-cutting enhancement layer over all six tasks rather than a seventh task.","The challenge sections (sparse, uncertain, error-prone, untrustworthy, and LLM-hybrid reasoning) define the open problems the next generation of KGR models must address."],"supporting_citations":[{"why":"Provides the closest prior scope by reviewing static, dynamic, and multi-modal KGR; the survey positions itself as extending this coverage.","marker":"[33]"},{"why":"Surveys few-shot and inductive KGR from the unseen-elements perspective, one of the six task families folded in.","marker":"[29]"},{"why":"Surveys few-shot KGR with structural and commonsense knowledge, covering the few-shot family.","marker":"[30]"},{"why":"Reviews inductive, few-shot, and zero-shot link prediction, covering the generalization-focused families.","marker":"[31]"},{"why":"Surveys knowledge graph embedding from the representation-space perspective, covering static single-step KGR.","marker":"[23]"},{"why":"Surveys the integration of logics and embeddings, covering the rule-based branch of static single-step KGR.","marker":"[25]"},{"why":"Surveys temporal KG representation learning and applications, covering the dynamic family.","marker":"[27]"},{"why":"Surveys temporal KG completion with a taxonomy, covering the dynamic family's temporal branch.","marker":"[28]"}],"fun_headline_variants":["Six reasoning tasks unite knowledge graph research","Unifying six knowledge graph reasoning tasks in one survey","Task-oriented taxonomy: the key to knowledge graph reasoning","From static to inductive: a full map of KG reasoning","KG reasoning mapped: six tasks, downstream uses, and LLMs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The taxonomy assumes that KGR research splits naturally and exhaustively into static single-step, static multi-step, dynamic, multi-modal, few-shot, and inductive reasoning; if a substantial body of work fits none of these six boxes, the survey's promise of comprehensiveness loses force.","fun_headline_variants_meta":{"raw":{"variants":["Six reasoning tasks unite knowledge graph research","Unifying six knowledge graph reasoning tasks in one survey","Task-oriented taxonomy: the key to knowledge graph reasoning","From static to inductive: a full map of KG reasoning","KG reasoning mapped: six tasks, downstream uses, and LLMs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000941,"raw_usage":{"total_tokens":4034,"prompt_tokens":972,"completion_tokens":3062,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":2985}},"tokens_in":588,"tokens_out":3062,"duration_ms":21208,"temperature":1.0,"reasoning_tokens":2985,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:58:50.694526+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the KGR papers published in the last three years, assign each to exactly one of the six task categories, and require that every assignment be unambiguous; if a sizable residue falls outside all six categories (e.g., hyper-relational or commonsense-only KGs, or zero-shot without a few-shot support set), the exhaustiveness claim is refuted.","supporting_citations":[{"cited_title":"A Survey On Few-shot Knowledge Graph Completion with Structural and Commonsense Knowledge","cited_arxiv_id":"2301.01172","evidence_quote":"Surveys few-shot KGR with structural and commonsense knowledge, covering the few-shot family."}],"review_version":1}