Pith. sign in

REVIEW 4 major objections 5 minor 45 references

Explicit temporal modelling and multi-branch evidence fusion are essential for faithful, multi-hop answers over historical case narratives.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-12 20:22 UTC pith:EKMVGIQD

load-bearing objection Solid systems integration of query-conditioned temporal GNN retrieval + multi-branch fusion that works on a custom case graph, but the leap to “essential for knowledge-grounded corpora” is not yet stress-tested outside that setup. the 4 major comments →

arxiv 2606.00029 v1 pith:EKMVGIQD submitted 2026-04-15 cs.CL cs.AI

TCAR-Gen: Temporal Graph Retrieval with Evidence Fusion for Knowledge-Grounded Generation

Sidra Nasir , Muhammad Noman Zahid , Rizwan Ahmed Khan This is my paper
classification cs.CL cs.AI
keywords LLMsKnowledge GraphsGraph Neural NetworksRetrieval-Augmented GenerationTemporal ReasoningEvidence FusionDomain-specific InformationExplainable AI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Standard retrieval-augmented generation often fails on complex questions about historical criminal cases because it matches isolated text chunks and cannot keep events in chronological order or fuse evidence from multiple sources. This paper introduces TCAR-Gen, a single pipeline that first builds a query-specific subgraph of entities, events and narrative chunks, scores those chunks with a mix of semantic similarity, graph relevance and temporal alignment, then explores several evidence-grounded reasoning branches (a Chain-of-Trees) before fusing the best paths into an answer. On the Victorian Crime Diaries benchmark of 160 gold-annotated queries spanning narrative, multi-hop, counterfactual and temporal types, the system reaches 0.3738 Recall@5 and perfect temporal consistency, far above four strong baselines. Ablations show that removing the context graph, query-conditioned attention or the temporal path penalty collapses performance; scaling tests show that the retrieval structure stays useful even for small language models, while generation quality still needs capacity. A sympathetic reader cares because the work supplies a concrete recipe for turning static case corpora into temporally coherent, verifiable answers rather than fluent but unsupported prose.

Core claim

The authors claim that treating retrieval and generation as one temporally grounded, graph-structured reasoning process—via a query-conditioned context graph, hybrid semantic–graph–temporal scoring, and multi-branch Chain-of-Trees fusion—produces substantially higher evidence recall and answer faithfulness than text-only, temporally re-ranked, or community-summary graph RAG on complex historical case questions.

What carries the argument

TCAR-Gen: a six-stage pipeline whose central mechanism is a query-conditioned temporal graph neural network that induces a compact context subgraph, supplies node relevance scores β for hybrid retrieval, and feeds multi-branch Chain-of-Trees reasoning whose paths are scored for evidence support, temporal consistency, graph coherence and model confidence.

Load-bearing premise

The whole comparison rests on one author-curated Victorian case graph and a 160-query gold set in which the graph neural network is trained on only twenty queries and every baseline scores near zero on multi-hop retrieval.

What would settle it

Re-run the identical seven query types and gold chunk annotations on a second, independently constructed temporal case corpus (or an open multi-hop temporal QA set) and check whether TCAR-Gen still lifts Recall@5 by roughly 0.3 over the same four baselines while preserving perfect temporal consistency.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes TCAR-Gen, a retrieval-augmented generation pipeline for temporally structured case narratives. It builds a query-specific context subgraph from a Neo4j knowledge graph, encodes it with a temporal GNN and query-conditioned attention pooling (QCAP), ranks narrative chunks with a hybrid semantic–graph–temporal score, and generates answers via multi-branch Chain-of-Trees reasoning with path scoring that includes a temporal-violation penalty. On an author-constructed Victorian Crime Diaries (VCD) benchmark of 160 stratified queries (20 train / 20 dev / 120 test), TCAR-Gen reports Recall@5 of 0.3738 and higher faithfulness and temporal consistency than Vanilla RAG, Temporal RAG, GraphRAG-C, and GraphRAG-T across seven query types. Eleven ablations and a five-model scaling study (GPT-OSS 20B to TinyLlama 1.1B) are used to argue that the context graph, query conditioning, and temporal penalty are critical, and that retrieval remains partly structural as model size shrinks while generation quality degrades.

Significance. If the gains hold under stronger multi-hop baselines and broader corpora, the work would be a useful systems contribution: it unifies document-level context, query-conditioned temporal graph encoding, hybrid retrieval, and multi-branch evidence fusion in one pipeline, with unusually thorough ablations and an explicit scaling study. The methodology is specified in enough detail (weights, hop budgets, path scoring, prompt template, training protocol) to be reproducible in principle. The main significance is therefore empirical and architectural rather than theoretical: a concrete demonstration that coarse community-summary GraphRAG and flat dense RAG are insufficient for multi-hop, temporally ordered forensic narratives when evidence is chunk-linked and chronologically constrained. That demonstration is currently limited by a single custom corpus and weak competitors, so the broader claim that temporal multi-branch fusion is “essential for knowledge-grounded corpora” is not yet established.

major comments (4)
  1. [§4.3 Baselines; Table 1; §5.2] Sections 4.1–4.3 and Table 1: the central superiority claim rests on four inference-only baselines whose absolute Recall@5 is ≤0.074 and 0.000 on Multi-Hop. GraphRAG-C/T produce community-level summaries yet are scored against chunk-level gold IDs of the form <case_id>_E01_C<idx>; by construction they cannot surface individual evidence chunks. Vanilla/Temporal RAG have no multi-hop expansion. The paper defers G-Retriever and RDPG for engineering reasons. Without at least one competitive multi-hop graph retriever adapted to the same schema (or an explicit chunk-level GraphRAG variant), the large margins do not yet show that TCAR-Gen’s temporal multi-branch design is necessary rather than that the chosen baselines are mismatched to the metric.
  2. [§4.2 Benchmark; §4.4 Implementation; §5.1] Sections 4.1–4.2: the GNN is supervised with BCE on only 20 train queries (gold_evidence positives vs other subgraph chunks), with hybrid and path weights grid-searched on 20 dev queries. Absolute Recall@5 of 0.3738 on 120 test queries is therefore estimated under very light structural supervision on an author-built graph and author-reviewed gold set. The manuscript should report variance (e.g., bootstrap CIs or multiple seeds), sensitivity to the 20-query train split, and a clearer statement that gold construction and schema design were independent of TCAR-Gen’s scoring function—beyond the brief assurance in §4.2—so that readers can judge overfitting risk to VCD structure.
  3. [Abstract; §6 Conclusion; §4.1] Abstract and §6: the conclusion that “explicit temporal modelling and multi-branch evidence fusion are essential for … knowledge-grounded corpora” overgeneralizes from a single domain corpus that the authors themselves flag as a scope constraint. Ablations A2/A5/A11 show large drops inside TCAR-Gen on VCD, but they do not test necessity outside this setup. The abstract and conclusion should be revised to claims scoped to temporally annotated multi-case narratives (or VCD-like graphs), or the evaluation should be extended to at least one external temporal multi-hop QA benchmark.
  4. [§3.3.3; Eq. (7), (10), (17); §5.2 Cross-Case] Eq. (7) and Eq. (10): temporal alignment is Jaccard overlap of intervals (or coarse narrative order when timestamps are missing). Eq. (17) applies a linear penalty η|V_temp(p)|. For a paper whose main selling point is temporal reasoning, this is a thin model of chronology (no ordering constraints, no continuous-time attention at retrieval time beyond time2vec in the GNN). The manuscript should either justify why interval Jaccard plus a post-hoc violation count is adequate for VCD, or report a failure analysis of temporal errors that this design cannot catch (e.g., same-interval events, missing dates, cross-case aliasing noted in §5.2).
minor comments (5)
  1. [§5.1; Figure 4] Figure 4 is referenced for comparative metrics but the manuscript text does not tabulate the full numeric comparison (only Recall@5 and selected generation metrics appear in prose). A single results table with all metrics for all systems would improve readability.
  2. [§3.5–3.6; §4.4] Notation inconsistency: the abstract and intro use “chain-of-trees” while §3.5 capitalizes “Chain-of-Trees”; λ and w weights are introduced without a single hyperparameter table. A compact config table would help.
  3. [Abstract; §4.2–4.3; §3.6.4] Typographical issues: “independently ofquery”, “failto integrate”, “T emporal”, “V anilla”, and run-together text in §3.6.4 (“subjecttoaminimum…”) should be cleaned in copy-editing.
  4. [§2 Related Work] Related work is thorough but long; a short positioning paragraph after Figure 1 already exists—consider trimming redundant temporal-KG citations that are not used as baselines.
  5. [§4.4 Evaluation Metrics; §5.1] Answer Relevancy uses the same MiniLM embeddings as retrieval; note potential circularity with semantic retrieval when interpreting that metric, even though the gap vs GraphRAG-C is negligible.

Circularity Check

0 steps flagged

No circularity: empirical systems paper with standard hyperparameter tuning and supervised GNN; no derivation reduces to its inputs by construction.

full rationale

TCAR-Gen is an empirical retrieval-augmented generation framework paper. Its central claims are measured performance numbers (Recall@5 = 0.3738, faithfulness, temporal consistency) on a held-out 120-query test split of a custom Victorian Crime Diaries Neo4j graph, plus ablation drops when components are removed. Hybrid retrieval weights (λs = 0.5, λg = 0.3, λt = 0.2) and path-scoring weights are obtained by ordinary grid search on a 20-query dev split and then frozen; the GNN is trained with binary cross-entropy on gold_evidence positives from a separate 20-query train split. These are standard supervised/tuning practices, not cases in which a fitted quantity is renamed a “prediction” of a closely related target. No equation equates an output to an input by definition, no uniqueness theorem is imported from the authors’ prior work, and no self-citation is load-bearing for the architecture. The authors themselves note the single-corpus scope limitation. Custom gold annotations and query generation from the same schema raise ordinary generalizability concerns, but they do not make any reported metric definitionally identical to the method’s inputs. Consequently the derivation chain (architecture → retrieval scores → generation metrics) contains no circular step under the defined patterns.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 3 invented entities

The central empirical claim rests on a custom temporally annotated crime KG, a small supervised GNN fit, hand-chosen hybrid and path-scoring weights, and modeling choices (bounded multi-hop expansion, three fixed branch types, temporal Jaccard alignment, RAGAS faithfulness). No new physical entities are postulated; the invented pieces are architectural modules. Free parameters and domain assumptions about the VCD schema carry most of the load.

free parameters (5)
  • hybrid retrieval weights λs, λg, λt = 0.5 / 0.3 / 0.2
    Set to 0.5, 0.3, 0.2 by grid search on the 20-query dev split; directly control Score(c|q) and thus Recall@K.
  • path scoring weights we, wt, wg, wm = 0.3 / 0.3 / 0.2 / 0.2
    Fixed at 0.3, 0.3, 0.2, 0.2 on dev; determine which Chain-of-Trees paths survive fusion and affect faithfulness/TC.
  • GNN training on 20-query train split = 20 queries, 50 epochs, lr=1e-3
    Binary CE on gold_evidence positives vs other subgraph chunks; small labeled set makes encoder fit highly dataset-specific.
  • temporal penalty coefficient η and path prune threshold S(p)<0.2 = η unspecified numeric; prune <0.2
    Control Temp(p) and branch pruning; chosen as part of the fusion design and held fixed.
  • max hop depth / node budget / top-K=10 / tree depth=3 = K=10; depth≤3; hop/budget stated as bounded
    Hard architectural caps that define Gctx(q) coverage and reasoning breadth; not learned from first principles.
axioms (5)
  • domain assumption Victorian Crime Diaries can be faithfully encoded as a heterogeneous Neo4j KG (Case/Person/Evidence/Location/Chunk) with reliable timestamps and typed edges sufficient for multi-hop forensic QA.
    Section 4.1 makes the corpus and schema the sole evaluation substrate; all temporal and graph claims depend on this encoding quality.
  • domain assumption Author-generated gold_evidence chunk IDs and hop counts correctly identify necessary evidence and are not trivially keyword-retrievable.
    Section 4.2: manual review of gold annotations; retrieval metrics are defined against these labels.
  • ad hoc to paper Community-summary GraphRAG and flat dense RAG with default configs are fair chunk-level retrieval baselines for this schema.
    Section 4.3; baselines hit 0.000 on Multi-Hop, which may partly reflect evaluation mismatch rather than pure capability.
  • ad hoc to paper Temporal alignment via interval Jaccard (or coarse narrative order) plus a linear temporal-violation penalty adequately captures chronological validity.
    Eqs. (10) and (17); standard engineering choice, not derived from a formal temporal logic.
  • standard math Standard message-passing GNN + attention math and cosine similarity are valid building blocks.
    Eqs. (3)–(9); conventional ML machinery.
invented entities (3)
  • TCAR-Gen pipeline (query-specific context graph + QCAP temporal GNN + hybrid score + Chain-of-Trees + path fusion) no independent evidence
    purpose: Unify contextual, relational, and temporal retrieval with multi-branch evidence-grounded generation.
    Architectural composite introduced in Section 3; evaluated only on VCD in this paper.
  • Chain-of-Trees reasoning with fixed branch types (witness, temporal overlap, shared evidence) no independent evidence
    purpose: Explore multiple evidence-grounded hypotheses before answer synthesis.
    Section 3.5; variant of multi-path CoT specialized to three forensic branch templates.
  • Victorian Crime Diaries (VCD) benchmark with seven query types and gold annotations no independent evidence
    purpose: Provide a temporally annotated multi-hop QA testbed for graph-aware RAG.
    Section 4.2; purpose-built by the authors; not an established public leaderboard in the paper.

pith-pipeline@v1.1.0-grok45 · 22426 in / 4158 out tokens · 39358 ms · 2026-07-12T20:22:43.598372+00:00 · methodology

0 comments
read the original abstract

Retrieval-augmented generation systems struggle with temporal reasoning and evidence fusion when answering complex questions over historical criminal case narratives. Existing approaches either retrieve independently of query semantics or fail to integrate multiple evidence sources coherently. We propose Temporal Context Augmented Retrieval Generation (TCAR-Gen), a framework that combines query-conditioned graph neural networks, temporal evidence fusion, and chain-of-trees reasoning to ground answer generation in retrieved evidence. On the Victorian Crime Diaries benchmark, TCAR-Gen achieves 0.3738 Recall@5, outperforming Vanilla RAG, Temporal RAG, GraphRAG-C, and GraphRAG-T across seven query types including multi-hop reasoning and counterfactual questions. Ablation studies reveal that the context graph, temporal penalty mechanism, and query conditioning are critical components. Cross-model evaluation across five language model (GPT-OSS 20B to TinyLlama 1.1B) demonstrates that TCAR-Gen maintains robust retrieval coverage at smaller model scales, though generation quality degrades substantially with reduced model capacity. Our work shows that explicit temporal modelling and multi-branch evidence fusion are essential for faithful, reasoning-intensive question answering over knowledge-grounded corpora.

Figures

Figures reproduced from arXiv: 2606.00029 by Muhammad Noman Zahid, Rizwan Ahmed Khan, Sidra Nasir.

Figure 1
Figure 1. Figure 1: Schematic overview of prior work—covering text-based RAG, graph-based reasoning, tem [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed Temporal Context Augmented Retrieval Generation (TCAR-Gen) [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Global temporal knowledge graph and query-specific subgraph with query-conditioned at [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparative evaluation of TCAR-Gen and four baselines on the Victorian Crime Diaries [PITH_FULL_IMAGE:figures/full_fig_p013_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Radar Chart: Recall@5 by query type across five LLM models. [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 9 linked inside Pith

  1. [1]

    Is chatgpt a general-purpose natural language processing task solver?,

    C. Qin, A. Zhang, Z. Zhang, J. Chen, M. Yasunaga, and D. Yang, “Is chatgpt a general-purpose natural language processing task solver?,”arXiv preprint arXiv:2302.06476, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Check your facts and try again: Improving large language models with external knowledge and automated feedback,

    B. Peng, M. Galley, P. He, H. Cheng, Y. Xie, Y. Hu, Q. Huang, L. Liden, Z. Yu, W. Chen,et al., “Check your facts and try again: Improving large language models with external knowledge and automated feedback,”arXiv preprint arXiv:2302.12813, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Improving language models by retrieving from trillions of tokens,

    S. Borgeaud, A. Mensch, J. Hoffmann, T. Cai, E. Rutherford, K. Millican, G. B. Van Den Driessche, J.-B. Lespiau, B. Damoc, A. Clark,et al., “Improving language models by retrieving from trillions of tokens,” inInternational conference on machine learning, pp. 2206–2240, PMLR, 2022

    2022

  4. [4]

    In-context retrieval-augmented language models,

    O. Ram, Y. Levine, I. Dalmedigos, D. Muhlgay, D. Jannai,et al., “In-context retrieval-augmented language models,”Transactions of the Association for Computational Linguistics, vol. 11, pp. 1316–1331, 2023

    2023

  5. [5]

    Kagnet: Knowledge-aware graph networks for commonsense reasoning,

    B. Y. Lin, X. Chen, J. Chen, and X. Ren, “Kagnet: Knowledge-aware graph networks for commonsense reasoning,”arXiv preprint arXiv:1909.02151, 2019

    Pith/arXiv arXiv 1909

  6. [6]

    Scalable multi-hop relational reasoning for knowledge-aware question answering,

    Y. Feng, X. Chen, B. Y. Lin, P. Wang, J. Li, and X. Ren, “Scalable multi-hop relational reasoning for knowledge-aware question answering,” inProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1295–1309, 2020

    2020

  7. [7]

    A survey of graph meets large language model: Progress and future directions,

    Y. Li, Z. Li, P. Wang, J. Li, X. Sun, H. Cheng, and J. X. Yu, “A survey of graph meets large language model: Progress and future directions,”arXiv preprint arXiv:2311.12399, 2023

    Pith/arXiv arXiv 2023

  8. [8]

    Retrieval-augmented generation for knowledge-intensive nlp tasks,

    P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W.-t. Yih, T. Rocktäschel, S. Riedel, and D. Kiela, “Retrieval-augmented generation for knowledge-intensive nlp tasks,” inAdvances in Neural Information Processing Systems, 2020

    2020

  9. [9]

    Context-alignment: Activating and enhancing llm capa- bilities in time series,

    Y. Hu, Q. Li, D. Zhang, J. Yan, and Y. Chen, “Context-alignment: Activating and enhancing llm capa- bilities in time series,”arXiv preprint arXiv:2501.03747, 2025

    arXiv 2025

  10. [10]

    Retrieval in decoder benefits generative models for explainable complex question answering,

    J. Feng, Q. Wang, H. Qiu, and L. Liu, “Retrieval in decoder benefits generative models for explainable complex question answering,”Neural Networks, vol. 181, p. 106833, 2025

    2025

  11. [11]

    Open domain question answering using early fusion of knowledge bases and text,

    H. Sun, B. Dhingra, M. Zaheer, K. Mazaitis, R. Salakhutdinov, and W. W. Cohen, “Open domain question answering using early fusion of knowledge bases and text,” inProceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pp. 4231–4242, 2018

    2018

  12. [12]

    Pullnet: Open domain question answering with iterative retrieval on knowledge bases and text,

    H. Sun, T. Bedrax-Weiss, and W. W. Cohen, “Pullnet: Open domain question answering with iterative retrieval on knowledge bases and text,” inProceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, pp. 2380–2390, 2019

    2019

  13. [13]

    Give us the facts: Enhancing large language models with knowledge graphs for fact-aware language modeling,

    L. Yang, H. Chen, Z. Li, X. Ding, and X. Wu, “Give us the facts: Enhancing large language models with knowledge graphs for fact-aware language modeling,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 7, pp. 3091–3110, 2024

    2024

  14. [14]

    From local to global: A graph rag approach to query-focused summarization,

    D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, D. Metropolitansky, R. O. Ness, and J. Larson, “From local to global: A graph rag approach to query-focused summarization,”arXiv preprint arXiv:2404.16130, 2024

    Pith/arXiv arXiv 2024

  15. [15]

    G-retriever: Retrieval-augmented generation for textual graph understanding and question answering,

    X. He, Y. Tian, Y. Sun, N. V. Chawla, T. Laurent, Y. LeCun, X. Bresson, and B. Hooi, “G-retriever: Retrieval-augmented generation for textual graph understanding and question answering,” inAdvances in Neural Information Processing Systems, 2024

    2024

  16. [16]

    Knowledge graph-extended retrieval augmented generation for question answering,

    J. Linders and J. M. Tomczak, “Knowledge graph-extended retrieval augmented generation for question answering,”Applied Intelligence, vol. 55, no. 17, p. 1102, 2025

    2025

  17. [17]

    Enhancing graph multi-hop reasoning for question answering with LLMs: An approach based on adaptive path generation,

    L. Ding, N. Ding, Q. Tao,et al., “Enhancing graph multi-hop reasoning for question answering with LLMs: An approach based on adaptive path generation,”Journal of Intelligent Information Systems, vol. 63, pp. 1455–1485, 2025

    2025

  18. [18]

    Rag-based explainable prediction of road users behaviors for automated driving using knowledge graphs and large language models,

    M. M. Hussien, A. N. Melo, A. L. Ballardini, C. S. Maldonado, R. Izquierdo, and M. Á. Sotelo, “Rag-based explainable prediction of road users behaviors for automated driving using knowledge graphs and large language models,”Expert Systems with Applications, vol. 265, p. 125914, 2025

    2025

  19. [19]

    Llm supporting knowledge tracing leveraging global subject and student specific knowledge graphs,

    L. Li, Z. Wang, J. M. Jose, and X. Ge, “Llm supporting knowledge tracing leveraging global subject and student specific knowledge graphs,”Information Fusion, p. 103577, 2025

    2025

  20. [20]

    Temporal knowledge graph ex- trapolation with subgraph information bottleneck,

    K. Chen, H. Yu, Y. Wang, X. Song, X. Zhao, Y. Xie, L. Gao, and A. Li, “Temporal knowledge graph ex- trapolation with subgraph information bottleneck,”Expert Systems with Applications, vol. 268, p. 126226, 2025. 18

    2025

  21. [21]

    Gs-kgc: A generative subgraph-based framework for knowledge graph completion with large language models,

    R. Yang, J. Zhu, J. Man, H. Liu, L. Fang, and Y. Zhou, “Gs-kgc: A generative subgraph-based framework for knowledge graph completion with large language models,”Information Fusion, vol. 117, p. 102868, 2025

    2025

  22. [22]

    Inductive representation learning on temporal graphs,

    D. Xu, C. Ruan, E. Korpeoglu, S. Kumar, and K. Achan, “Inductive representation learning on temporal graphs,”arXiv preprint arXiv:2002.07962, 2020

    Pith/arXiv arXiv 2002

  23. [23]

    Few-shot multi-hop reasoning via reinforcement learning and path search strategy over temporal knowledge graphs,

    L. Bai, H. Zhang, X. An, and L. Zhu, “Few-shot multi-hop reasoning via reinforcement learning and path search strategy over temporal knowledge graphs,”Information Processing & Management, vol. 62, no. 3, p. 104001, 2025

    2025

  24. [24]

    Spatiotemporal knowledge graph multi-hop reasoning based on large language models,

    X. Liang, X. Xu, R. Ma, L. Yan, and Z. Ma, “Spatiotemporal knowledge graph multi-hop reasoning based on large language models,”Engineering Applications of Artificial Intelligence, vol. 164, p. 113229, 2026

    2026

  25. [25]

    Temporal knowledge graph reasoning with dynamic memory enhancement,

    F. Zhang, Z. Zhang, F. Zhuang, Y. Zhao, D. Wang, and H. Zheng, “Temporal knowledge graph reasoning with dynamic memory enhancement,”IEEE Transactions on Knowledge and Data Engineering, vol. 36, no. 11, pp. 7115–7128, 2024

    2024

  26. [26]

    Temporal inductive path neural network for temporal knowledge graph reasoning,

    H. Dong, P. Wang, M. Xiao, Z. Ning, P. Wang, and Y. Zhou, “Temporal inductive path neural network for temporal knowledge graph reasoning,”Artificial Intelligence, vol. 329, p. 104085, 2024

    2024

  27. [27]

    Multi-hop path reasoning over sparse temporal knowledge graphs based on path completion and reward shaping,

    X. Meng, L. Bai, J. Hu, and L. Zhu, “Multi-hop path reasoning over sparse temporal knowledge graphs based on path completion and reward shaping,”Information Processing & Management, vol. 61, no. 2, p. 103605, 2024

    2024

  28. [28]

    Historical facts learning from long-short terms with language model for temporal knowledge graph reasoning,

    W. Xu, B. Liu, M. Peng, Z. Jiang, X. Jia, K. Liu, L. Liu, and M. Peng, “Historical facts learning from long-short terms with language model for temporal knowledge graph reasoning,”Information Processing & Management, vol. 62, no. 3, p. 104047, 2025

    2025

  29. [29]

    HGCT: Enhancing temporal knowledge graph reasoning through extrapolated historical fact extraction,

    H. Dao, N. Phan, T. Le, and N.-T. Nguyen, “HGCT: Enhancing temporal knowledge graph reasoning through extrapolated historical fact extraction,”Knowledge-Based Systems, vol. 316, p. 113358, 2025

    2025

  30. [30]

    Quadruple mention text-enhanced temporal knowledge graph reasoning,

    L. Zhu, W. Zhao, and L. Bai, “Quadruple mention text-enhanced temporal knowledge graph reasoning,” Engineering Applications of Artificial Intelligence, vol. 133, p. 108058, 2024

    2024

  31. [31]

    Re-segnn: recurrent semantic evidence-aware graph neural network for temporal knowledge graph forecasting,

    W. Cai, M. Li, X. Shi, Y. Fan, Q. Zhu, and H. Jin, “Re-segnn: recurrent semantic evidence-aware graph neural network for temporal knowledge graph forecasting,”Science China information sciences, vol. 68, no. 2, p. 122104, 2025

    2025

  32. [32]

    An inductive reasoning model based on interpretable logical rules over temporal knowledge graph,

    X. Mei, L. Yang, Z. Jiang, X. Cai, D. Gao, J. Han, and S. Pan, “An inductive reasoning model based on interpretable logical rules over temporal knowledge graph,”Neural Networks, vol. 174, p. 106219, 2024

    2024

  33. [33]

    Leveraging temporal validity of rules via llms for enhanced temporal knowledge graph reasoning,

    Q. Pan, L. Yao, G. Shen, X. Han, Y. Chen, and X. Kong, “Leveraging temporal validity of rules via llms for enhanced temporal knowledge graph reasoning,”Knowledge-Based Systems, p. 114094, 2025

    2025

  34. [34]

    Historical trends and normalizing flow for one-shot temporal knowledge graph reasoning,

    R. Ma, L. Wang, H. Wu, B. Gao, X. Wang, and L. Zhao, “Historical trends and normalizing flow for one-shot temporal knowledge graph reasoning,”Expert Systems With Applications, vol. 260, p. 125366, 2025

    2025

  35. [35]

    Multi-hop interpretable meta learning for few-shot temporal knowledge graph completion,

    L. Bai, S. Han, and L. Zhu, “Multi-hop interpretable meta learning for few-shot temporal knowledge graph completion,”Neural Networks, vol. 183, p. 106981, 2025

    2025

  36. [36]

    Arule-andquery-guidedreinforcementlearningforextrapolation reasoning in temporal knowledge graphs,

    T.Chen, L.Yang, Z.Wang, andJ.Long, “Arule-andquery-guidedreinforcementlearningforextrapolation reasoning in temporal knowledge graphs,”Neural Networks, vol. 185, p. 107186, 2025

    2025

  37. [37]

    Explainable reasoning over temporal knowledge graphs by pre-trained language model,

    Q. Li and G. Wu, “Explainable reasoning over temporal knowledge graphs by pre-trained language model,” Information Processing & Management, vol. 62, no. 1, p. 103903, 2025

    2025

  38. [38]

    Tempqa: An llm-based framework for temporal knowledge graph question answering,

    Q. Hu, X. Tu, A. Li, and B. Yao, “Tempqa: An llm-based framework for temporal knowledge graph question answering,”Knowledge-Based Systems, p. 114988, 2025

    2025

  39. [39]

    Semantic-enhanced reasoning question answering over temporal knowledge graphs,

    C. Du, X. Li, and Z. Li, “Semantic-enhanced reasoning question answering over temporal knowledge graphs,”Journal of Intelligent Information Systems, vol. 62, pp. 859–881, 2024

    2024

  40. [40]

    TEQA: Temporal knowledge graph enhanced question answering,

    Q. Liu, S. Feng, and M. Huang, “TEQA: Temporal knowledge graph enhanced question answering,” Knowledge-Based Systems, vol. 325, p. 113916, 2025

    2025

  41. [41]

    Chain- of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V. Le, and D. Zhou, “Chain- of-thought prompting elicits reasoning in large language models,”arXiv preprint arXiv:2201.11903, 2022

    Pith/arXiv arXiv 2022

  42. [42]

    Self- consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self- consistency improves chain of thought reasoning in language models,”arXiv preprint arXiv:2203.11171, 2022. 19

    Pith/arXiv arXiv 2022

  43. [43]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” inPro- ceedings of the 2019 conference on empirical methods in natural language processing and the 9th interna- tional joint conference on natural language processing (EMNLP-IJCNLP), pp. 3982–3992, 2019

    2019

  44. [44]

    Ragas: Automated evaluation of retrieval augmented gen- eration,

    S. Es, J. James, L. E. Anke, and S. Schockaert, “Ragas: Automated evaluation of retrieval augmented gen- eration,” inProceedings of the 18th conference of the european chapter of the association for computational linguistics: system demonstrations, pp. 150–158, 2024

    2024

  45. [45]

    Time2vec: Learning a vector representation of time,

    S. M. Kazemi and D. Poole, “Time2vec: Learning a vector representation of time,”arXiv preprint arXiv:1907.05321, 2019. 20

    Pith/arXiv arXiv 1907