Pith. sign in

REVIEW 4 major objections 8 minor 68 references

Are Large Language Models Good Temporal Graph Learners?

T0 review · 4 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read TGTalker, a training-free framework, converts recent edges of a temporal network into text and shows that pre-trained LLMs predict future links competitively with specialized temporal graph neural networks, consistently beating TGN and…

desk verdict Useful and timely framework for LLMs on real-world temporal graphs, but Table 3's MRR numbers are not verifiable until the authors specify how a single LLM prediction is converted to a ranked list. read the letter →

arxiv 2506.05393 v1 pith:EKIKWWAJ submitted 2025-06-03 cs.CL cs.LG

Shenyang Huang , Ali Parviz , Emma Kondrup , Zachary Yang , Zifeng Ding , Michael Bronstein , Reihaneh Rabbany , Guillaume Rabusseau This is my paper
classification cs.CLcs.LG
keywords temporalgraphlearninglargelanguagemodelsin-contextlinkpredictionrecencybiasneighborsamplingexplainabilityTGTalker
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces TGTalker, a framework that turns a slice of a temporal network—the most recent edges, a few in-context examples, and the source node's recent neighbors—into a text prompt for a pre-trained large language model. The central claim is that, with no training or fine-tuning, LLMs prompted this way perform link prediction competitively with state-of-the-art temporal graph neural networks, outperforming popular models such as TGN and HTGN on all five real-world networks tested. If true, this means the hard part of temporal link prediction can sometimes be reduced to deciding which recent structure to show the model, and that LLMs can serve as plug-in, explainable predictors for evolving graphs.

What carries the argument

The central mechanism is the prompt composition rule of TGTalker, built from a background set (the b most recent edges as text), an example set (five-shot in-context learning), a query set that encodes the source node and timestamp, and temporal neighbor sampling that injects the m most recent one-hop neighbors of the source. The load-bearing idea is the recency bias of temporal graphs, which justifies compressing the full history into this small window; the ablation shows that removing temporal neighbors collapses MRR from 0.649 to 0.322 on tgbl-wiki, confirming that this component, not the LLM's priors, carries most of the prediction signal.

What would settle it

Construct a temporal graph (or filter an existing one) whose next link is determined by an interaction older than the background window—say, a periodic edge that reappears every 1,000 steps but is absent from the last 300. If TGTalker ranks that true destination far worse than TGN or TNCN on those edges while matching them on recency-driven edges, the recency-bias assumption is the reason.

Watch

Extended reading notes

Core claim

TGTalker's claim is that a pre-trained LLM, given a prompt built from (a) the b most recent edges before the query timestamp, (b) five in-context question-answer examples, and (c) the m most recent one-hop neighbors of the source node, can directly name the most likely destination node of a temporal link. Across tgbl-wiki, Reddit, LastFM, UCI, and Enron, the best TGTalker variants reach test MRR of 0.651, 0.626, 0.079, 0.220, and 0.200 respectively, ranking in the top three among all compared methods on every dataset and consistently beating TGN and HTGN. The paper further claims that the same LLM can produce natural-language explanations for each prediction, and that these explanations cluster into ten categories that mirror known temporal graph heuristics such as 'most recent interaction' and 'most common destination'.

Load-bearing premise

The decisive evidence for a prediction must be among the b most recent edges or the m most recent neighbors of the source node; links that depend on older history are invisible to the prompt, and the paper's own appendix concedes this may hurt on datasets with long-range dependencies.

Editorial extensions

If this is right

  • Temporal link prediction can be done without training a model on the dataset: a practitioner can adapt TGTalker to a new network by rewriting a prompt, then swap in a better LLM later.
  • LLM-generated explanations assign each prediction to one of ten reasoning categories, meaning temporal link prediction becomes auditable in natural language rather than a bare probability.
  • The consistent win over TGN and HTGN, which see the entire history, indicates that recent structure plus in-context examples is often sufficient for these five networks.
  • Open-weight models (Qwen2.5-7B) can beat a larger closed model (GPT-4.1-mini) on several datasets, so the method's ceiling depends on instruction-tuning quality, not parameter count.
  • Because TGTalker uses a fixed-size window, its cost per prediction is bounded by the LLM context length, so the approach remains usable on graphs whose full edge count far exceeds the model's maximum input size.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the prompt is dominated by the most recent edges and neighbors, TGTalker is effectively an LLM-parameterized recency heuristic; the paper's own ablation leaves open whether the LLM contributes anything beyond executing that heuristic, which a direct comparison against a tuned EdgeBank variant would settle.
  • A natural stress test is long-range periodicity: on a dataset where the correct destination last appeared outside the background window, TGTalker should systematically fail, revealing the boundary of the recency assumption.
  • The ten explanation categories amount to a discovered taxonomy of temporal link heuristics; one could verify it by checking whether LLM explanations flagged as 'Most Recent Interaction' correspond exactly to EdgeBank's predictions on the same queries.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 8 minor

Summary. This paper introduces TGTalker, a training-free framework that uses pre-trained LLMs for temporal link prediction and explanation on continuous-time dynamic graphs. Temporal graph structure near a query is subsampled via a recent-edge background set, one-hop temporal neighbor sampling, and a few-shot example set, then verbalized into integer-encoded prompts. The paper evaluates six LLM variants on five datasets (tgbl-wiki, Reddit, LastFM, UCI, Enron) against TGNN baselines, reporting test MRR in Table 3: the best TGTalker variants rank second to TNCN on three datasets and third on two, while beating TGN and HTGN on all five. It also elicits textual explanations and clusters them into ten categories. The paper claims competitive performance with state-of-the-art TGNNs, consistent superiority over TGN/HTGN, training-free adaptability, and a new explainability avenue.

Significance. If the findings hold, TGTalker is a notable data point: a zero-training LLM prompting framework can rival trained TGNNs on real temporal graphs and additionally provide natural-language explanations. The paper's strengths include public code, evaluation across multiple LLM families and datasets, ablations of the main components, and an explicit acknowledgement of context-window and recency-bias limitations in Appendix A. However, the central comparison is currently under-specified because a single decoded node is not an MRR ranking, and the LLM results lack variance or significance testing. The explanation taxonomy is also generated and labeled by the same LLM family that is being analyzed, making the interpretability claims partly circular. These issues are fixable and do not appear to require a fundamentally different study, but they must be resolved before the headline claims can be taken as established.

major comments (4)
  1. [§4.1 and Table 3] Section 4.1 states that 'the LLM directly outputs the destination node' and that TGTalker avoids scoring all node pairs, yet Table 3 reports MRR, which is a ranking-based metric. The manuscript never specifies how a single emitted destination is converted to a rank: whether the TGB negative samples are included in the prompt, how an incorrect or malformed answer is scored, and what convention maps a miss to a rank position. Without this protocol, the reported MRR values are not reproducible and cannot be compared with the TGNN baselines, which rank all negative candidates. Please state the exact conversion rule and align it with TGB's official evaluation, or re-report with a metric that matches the actual output protocol.
  2. [§5.1, Table 3] The LLM results are reported from a single inference run (as noted in the Table 3 caption), while each TGNN baseline reports mean and standard deviation over five random seeds. The claims that TGTalker 'consistently outperforms' TGN and HTGN and is competitive with TNCN rest on point estimates with no significance testing, and stochastic decoding, prompt ordering, or decoding temperature could alter the ranks. Please report multiple runs with variance and significance tests, or justify determinism and provide a stability analysis showing that the qualitative conclusions are robust.
  3. [§5.2] The ten explanation categories are proposed by the same LLM family (Qwen3-8B and GPT-4.1-mini) that later generates the explanations and classifies them into the taxonomy. This creates a circular dependency: the category distribution in Figure 3 may reflect the model's self-consistency and prompt compliance rather than an independent characterization of its reasoning diversity. Please use predefined categories from human experts, or have independent human annotators label a sample with reported inter-annotator agreement, before drawing conclusions about the models' reasoning patterns.
  4. [Appendix G, Tables 8 and 9] The ablations show that removing temporal neighbor sampling drops tgbl-wiki MRR from 0.649 to 0.322 and Reddit MRR from 0.613 to 0.122, and the no-components configuration nearly destroys performance (0.008 and 0.002). This is strong evidence that the retrieval component carries most of the predictive signal. The paper should therefore qualify its title and interpretive claims: the framework demonstrates that LLMs can use retrieved temporal context effectively, but the results do not establish that LLMs are intrinsically strong temporal graph reasoners independent of the recency-based retrieval mechanism.
minor comments (8)
  1. [§2] The phrase 'graph-based pro3mpting' contains a typo and should read 'prompting'.
  2. [Figure 2] The figure label repeats 'Temporal Neighbors Temporal Neighbors'; one copy appears to be an error.
  3. [Appendix D] Appendix D mentions experiments with GPT-4o-mini, but this model does not appear in the model list or in Table 3; please clarify whether it was used.
  4. [Appendix G, Table 8] The heading 'Qwen3.8b' is inconsistent with the model name 'Qwen3-8B' used elsewhere in the paper.
  5. [Abstract and Appendix F] The abstract gives a GitHub URL while Appendix F gives an anonymous.4open.science URL; please state which repository is canonical for the published version.
  6. [References] The reference list contains duplicate entries for Rossi et al. 2020 ([45] and [46]); they should be merged.
  7. [Table 2] The column header 'Surprise Duration' is unclear; the column likely reports the dataset duration, and the header should be split or renamed.
  8. [§5.1] The text says 'five families of LLMs' but lists six models (two Qwen3 sizes, Qwen2.5, Mistral, Llama3, GPT-4.1-mini); please clarify the counting.

Circularity Check

1 steps flagged · score 2.0 of 10

Central MRR claim is self-contained against external benchmarks; only a mild self-referential loop in the explanation-taxonomy study prevents a zero score.

  1. self definitional [Section 5.2 (Temporal Link Explanation Results, subsection Designing Explanation Categories and Quantitative Analysis)]
    "We first use Qwen3-8B and GPT-4.1-mini to generate example link explanations, and we also ask them to design appropriate categories for explanations. Then we retrieve the categories and perform human annotation to ensure a set of ten reasonable and comprehensive categories. ... The same LLM is used for prediction and explanation classification because only the prediction LLM knows the reasoning for its prediction."

    The explanation analysis is closed by construction: the LLM families under study propose the ten-category taxonomy, generate the explanations, and then label those explanations using the self-generated taxonomy. Therefore the reported category distributions, and the claim that LLMs 'naturally discover' reasoning patterns analogous to EdgeBank or PopTrack, are partly guaranteed by the measurement design rather than by independent evidence. Human annotation curates the LLM-proposed categories but does not supply an independent category source, so the loop remains. This self-confirmation affects only the explanation study; it does not touch the Table 3 link prediction evaluation, which uses fixed held-out edges and no fitted parameters.

full rationale

The paper's central link prediction claim is not circular: TGTalker has no trainable parameters, is not fitted to the test set, and Table 3 reports MRR against ground-truth temporal edges with the fixed TGB negative-sampling procedure, an external benchmark. The recency-bias subsampling is an explicit modeling assumption, and the ablation in Appendix G shows performance collapses without temporal-neighbor context, so the method's output is not identical to its input by construction. The skeptic's concern about converting a single LLM-emitted destination into an MRR rank is a transparency and reproducibility issue, not a circularity step, because no passage in the paper equates the emitted node with a rank by construction. The only genuine self-referential element is the explanation-taxonomy study, where the same LLM family proposes the categories, writes the explanations, and classifies those explanations; this mild measurement loop keeps the score at 2 rather than 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on four hand-chosen hyperparameters and four domain assumptions about temporal graph structure and LLM behavior. No new physical or mathematical entities are introduced. The recency bias assumption is the most consequential because it drives the context selection mechanism.

free parameters (4)
  • background_set_size = 300
    Number of most recent edges included in the background set; chosen by hand in Section 5 as default (b=300).
  • temporal_neighbor_count = 2
    Number of 1-hop temporal neighbors sampled per source node; default m=2 in Section 5, ablation in Appendix G shows strong performance effect.
  • example_set_size = 5
    Number of in-context learning examples; set to 5 in Section 4.2 and 5.
  • batch_size = 200
    Test edges per prediction batch; fixed at 200 in Section 5.
assumptions (4)
  • domain assumption Strong recency bias in temporal graphs
    Invoked in Section 4.2 to justify keeping only the b most recent edges and m recent neighbors. The method fails if relevant predictive signals are old.
  • domain assumption Pre-trained LLMs can perform few-shot in-context learning on verbalized graph tuples
    The entire framework assumes that instruction-tuned LLMs will follow the prompt format and infer link patterns from integer node IDs and timestamps.
  • domain assumption Textual tuple encoding preserves enough structure for link prediction
    TGTalker encodes edges as (src, dst, timestamp) tuples; this assumes no loss of critical relational information compared to vector representations.
  • domain assumption Streaming evaluation with no gradient updates at test time
    Section 3 defines the standard CTDG streaming setting; TGTalker's use of test-time neighbor sampling depends on this protocol being valid for comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Are Large Language Models Good Temporal Graph Learners?." pith.science (2026). https://pith.science/paper/EKIKWWAJ

@misc{pith2026250605393,
  author       = {Pith},
  title        = {Pith review of: Are Large Language Models Good Temporal Graph Learners?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EKIKWWAJ}},
  note         = {Machine review of arXiv:2506.05393}
}
read the original abstract

Large Language Models (LLMs) have recently driven significant advancements in Natural Language Processing and various other applications. While a broad range of literature has explored the graph-reasoning capabilities of LLMs, including their use of predictors on graphs, the application of LLMs to dynamic graphs -- real world evolving networks -- remains relatively unexplored. Recent work studies synthetic temporal graphs generated by random graph models, but applying LLMs to real-world temporal graphs remains an open question. To address this gap, we introduce Temporal Graph Talker (TGTalker), a novel temporal graph learning framework designed for LLMs. TGTalker utilizes the recency bias in temporal graphs to extract relevant structural information, converted to natural language for LLMs, while leveraging temporal neighbors as additional information for prediction. TGTalker demonstrates competitive link prediction capabilities compared to existing Temporal Graph Neural Network (TGNN) models. Across five real-world networks, TGTalker performs competitively with state-of-the-art temporal graph methods while consistently outperforming popular models such as TGN and HTGN. Furthermore, TGTalker generates textual explanations for each prediction, thus opening up exciting new directions in explainability and interpretability for temporal link prediction. The code is publicly available at https://github.com/shenyangHuang/TGTalker.

Figures

Figures reproduced from arXiv: 2506.05393 by the authors.

Figure 1
Figure 1. Overview of TGTalker framework. This flexibility in learning approaches has opened new avenues for applying LLMs to complex structured data. To harness these capabilities for graph-structured data, recent research has explored various ap￾proaches to integrate LLMs with graph reasoning tasks. This integration is particularly relevant given that many real-world systems—from molecular structures [2] and social networks… view at source ↗
Figure 2
Figure 2. TGTalker framework has four core components: a background set, an example set, a query [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Explanation category composition (top row) and test MRR per category (bottom row) plots [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Explanation category composition plots for [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 42 canonical work pages

  1. [2]

    Beaini, S

    D. Beaini, S. Huang, J. A. Cunha, Z. Li, G. Moisescu-Pareja, O. Dymov, S. Maddrell-Mander, C. McLean, F. Wenkel, L. Müller, et al. Towards foundational models for molecular learning on large-scale multi-task datasets. InICLR, 2024

    work page 2024

  2. [3]

    Behrouz, A

    A. Behrouz, A. Parviz, M. Karami, C. Sanford, B. Perozzi, and V . S. Mirrokni. Best of both worlds: Advantages of hybrid graph sequence models.ArXiv, abs/2411.15671, 2024

    arXiv 2024

  3. [4]

    Broder, R

    A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins, and J. Wiener. Graph structure in the web.Computer Networks, 33(1-6):309–320, 2000

    work page 2000

  4. [5]

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-V oss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. ...

    1901

  5. [6]

    Calzolari, A

    G. Calzolari, A. Ichino, M. Nardotto, and G. Zanella. Large language models as tax attorneys: a case study in legal analysis.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 382(2284):20230159, 2024

    work page 2024

  6. [7]

    D. S. Chaplot et al. Mistral 7b. https://arxiv.org/abs/2310.06825, 2023. Version 0.1, Mistral AI. For instruct fine-tuned version see https://huggingface.co/mistralai/ Mistral-7B-Instruct-v0.3

    arXiv 2023

  7. [8]

    J. Chen, X. Wang, and X. Xu. Gc-lstm: Graph convolution embedded lstm for dynamic network link prediction.Applied Intelligence, pages 1–16, 2022

    2022

  8. [9]

    Chen and R

    J. Chen and R. Ying. Tempme: Towards the explainability of temporal graph neural networks via motif discovery.Advances in Neural Information Processing Systems, 36:29005–29028, 2023

    work page 2023

Show all 68 references
  1. [10]

    Coda-Forno, M

    J. Coda-Forno, M. Binz, Z. Akata, M. Botvinick, J. X. Wang, and E. Schulz. Meta-in-context learning in large language models. InAdvances in Neural Information Processing Systems, 2023

    2023

  2. [11]

    W. Cong, S. Zhang, J. Kang, B. Yuan, H. Wu, X. Zhou, H. Tong, and M. Mahdavi. Do we really need complicated model architectures for temporal networks? InThe Eleventh International Conference on Learning Representations, 2022

    2022

  3. [12]

    Cornell, O

    F. Cornell, O. Smirnov, G. Z. Gandler, and L. Cao. On the power of heuristics in temporal graphs.arXiv preprint arXiv:2502.04910, 2025

    2025 arXiv

  4. [13]

    X. Dai, Q. Wen, Y . Shen, H. Wen, D. Li, J. Tang, and C. Shan. Revisiting the graph reasoning ability of large language models: Case studies in translation, connectivity and shortest path. arXiv preprint arXiv:2408.09529, 2024. 10

    2024 arXiv

  5. [14]

    Daniluk and J

    M. Daniluk and J. Dabrowski. Temporal graph models fail to capture global temporal dynamics. InTemporal Graph Learning Workshop@ NeurIPS 2023

    2023

  6. [15]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In J. Burstein, C. Doran, and T. Solorio, editors, Proceedings of the 2019 Conference of the North American Chapter of the Association for Comput...

    2019

  7. [16]

    Z. Ding, H. Cai, J. Wu, Y . Ma, R. Liao, B. Xiong, and V . Tresp. zrllm: Zero-shot relational learning on temporal knowledge graphs with large language models. In K. Duh, H. Gómez- Adorno, and S. Bethard, editors,Proceedings of the 2024 Conference of the North American Chapter...

    2024

  8. [17]

    H. Dong, Z. Zhang, Y . Li, Z. Li, Y . Wang, Y . Liu, W. Chen, et al. A survey on in-context learning.arXiv preprint arXiv:2301.00234, 2023

    2023 arXiv

  9. [18]

    Dubey, L

    A. Dubey, L. Van Der Maaten, et al. The llama 3 herd of models, 2024. Version 3

    2024

  10. [19]

    V . P. Dwivedi, L. Rampášek, M. Galkin, A. Parviz, G. Wolf, A. T. Luu, and D. Beaini. Long range graph benchmark.Advances in Neural Information Processing Systems, 35:22326–22340, 2022

    2022

  11. [20]

    Fatemi, J

    B. Fatemi, J. Halcrow, and B. Perozzi. Talk like a graph: Encoding graphs for large language models.arXiv preprint arXiv:2310.04560, 2023

    2023 arXiv

  12. [21]

    Fatemi, J

    B. Fatemi, J. Halcrow, and B. Perozzi. Talk like a graph: Encoding graphs for large language models. InInternational Conference on Learning Representations (ICLR), 2024

    2024

  13. [22]

    Franceschelli et al

    G. Franceschelli et al. On the creativity of large language models.arXiv preprint arXiv:2304.00008, 2025

    2025 arXiv

  14. [23]

    Gastinger, S

    J. Gastinger, S. Huang, M. Galkin, E. Loghmani, A. Parviz, F. Poursafaei, J. Danovitch, E. Rossi, I. Koutis, H. Stuckenschmidt, R. Rabbany, and G. Rabusseau. Tgb 2.0: A benchmark for learning on temporal knowledge graphs and heterogeneous graphs.Advances in Neural Information ...

    2024

  15. [24]

    W. L. Hamilton, R. Ying, and J. Leskovec. Inductive representation learning on large graphs. In Advances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  16. [25]

    Huang, H

    Q. Huang, H. Ren, P. Chen, G. Kržmanc, D. Zeng, P. S. Liang, and J. Leskovec. Prodigy: Enabling in-context learning over graphs.Advances in Neural Information Processing Systems, 36:16302–16317, 2023

    2023

  17. [26]

    Huang, F

    S. Huang, F. Poursafaei, J. Danovitch, M. Fey, W. Hu, E. Rossi, J. Leskovec, M. Bronstein, G. Rabusseau, and R. Rabbany. Temporal graph benchmark for machine learning on temporal graphs.Advances in Neural Information Processing Systems, 36:2056–2073, 2023

    2023

  18. [27]

    Huang, F

    S. Huang, F. Poursafaei, R. Rabbany, G. Rabusseau, and E. Rossi. Utg: Towards a unified view of snapshot and event based models for temporal graphs. InThe Third Learning on Graphs Conference

  19. [28]

    Z. Ji, J. Lee, T. Sun, Z. Du, Y . Zhang, Z. Lin, X. Deng, Y . Song, Y . Tan, Q. Zhou, et al. Siren’s song in the ai ocean: A survey on hallucination in large language models.arXiv preprint arXiv:2309.01219, 2023

    2023 arXiv

  20. [29]

    S. M. Kazemi, R. Goel, K. Jain, I. Kobyzev, A. Sethi, P. Forsyth, and P. Poupart. Representation learning for dynamic graphs: A survey.Journal of Machine Learning Research, 21(70):1–73, 2020. 11

    2020

  21. [30]

    T. N. Kipf and M. Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations, 2016

    2016

  22. [31]

    Kumar, X

    S. Kumar, X. Zhang, and J. Leskovec. Predicting dynamic embedding trajectory in temporal interaction networks. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pages 1269–1278, 2019

    2019

  23. [32]

    D. Lee, K. Ahrabian, W. Jin, F. Morstatter, and J. Pujara. Temporal knowledge graph forecasting without knowledge using in-context learning. In H. Bouamor, J. Pino, and K. Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EM...

    2023

  24. [33]

    J. Li, R. Wu, Y . Zhu, H. Zhang, L. Chen, and Z. Zheng. Are large language models in-context graph learners?arXiv preprint arXiv:2502.13562, 2025

    2025 arXiv

  25. [34]

    R. Liao, X. Jia, Y . Li, Y . Ma, and V . Tresp. Gentkg: Generative forecasting on temporal knowledge graph with large language models. In K. Duh, H. Gómez-Adorno, and S. Bethard, editors,Findings of the Association for Computational Linguistics: NAACL 2024, Mexico City, Mexico...

    2024

  26. [35]

    Limonad, F

    L. Limonad, F. Fournier, J. M. V . Díaz, I. Skarbovsky, S. Gur, and R. Lazcano. Monetizing currency pair sentiments through llm explainability. InEuropean Conference on Artificial Intelligence, 2024

    2024

  27. [36]

    H. Liu, C. Li, Q. Wu, and Y . J. Lee. Visual instruction tuning. InAdvances in Neural Information Processing Systems, volume 36, pages 26113–26129, 2023

    2023

  28. [37]

    Longa, V

    A. Longa, V . Lachi, G. Santin, M. Bianchini, B. Lepri, P. Liò, F. Scarselli, A. Passerini, et al. Graph neural networks for temporal graphs: State of the art, open challenges, and opportunities. TRANSACTIONS ON MACHINE LEARNING RESEARCH, 2023

    2023

  29. [38]

    Luo and P

    Y . Luo and P. Li. Neighborhood-aware scalable temporal network representation learning. In Learning on Graphs Conference, pages 1–1. PMLR, 2022

    2022

  30. [39]

    McKenna, T

    N. McKenna, T. Li, L. Cheng, M. J. Hosseini, M. Johnson, and M. Steedman. Sources of hallucination by large language models on inference tasks.arXiv preprint arXiv:2305.14552, 2023

    2023 arXiv

  31. [40]

    Nguyen, L

    M.-V . Nguyen, L. Luo, F. Shiri, D. Phung, Y .-F. Li, T.-T. Vu, and G. Haffari. Direct evaluation of chain-of-thought in multi-hop reasoning with knowledge graphs. In L.-W. Ku, A. Martins, and V . Srikumar, editors,Findings of the Association for Computational Linguistics: ACL...

    2024

  32. [41]

    Chatgpt (mar 14 version) [large language model]

    OpenAI. Chatgpt (mar 14 version) [large language model]. https://chat.openai.com/ chat, 2023

    2023

  33. [42]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    OpenAI et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    2023 arXiv

  34. [43]

    Perozzi, B

    B. Perozzi, B. Fatemi, D. Zelle, A. Tsitsulin, M. Kazemi, R. Al-Rfou, and J. J. Halcrow. Let your graph do the talking: Encoding structured data for llms.ArXiv, abs/2402.05862, 2024

    2024 arXiv

  35. [44]

    Poursafaei, S

    F. Poursafaei, S. Huang, K. Pelrine, and R. Rabbany. Towards better evaluation for dynamic link prediction.Advances in Neural Information Processing Systems, 35:32928–32941, 2022

    2022

  36. [46]

    Rossi, B

    E. Rossi, B. Chamberlain, F. Frasca, D. Eynard, F. Monti, and M. Bronstein. Temporal graph networks for deep learning on dynamic graphs.arXiv preprint arXiv:2006.10637, 2020

    2006 arXiv

  37. [47]

    Sanford, B

    C. Sanford, B. Fatemi, E. Hall, A. Tsitsulin, S. M. Kazemi, J. J. Halcrow, B. Perozzi, and V . S. Mirrokni. Understanding transformer reasoning capabilities via graph algorithms.ArXiv, abs/2405.18512, 2024. 12

    2024 arXiv

  38. [48]

    Schulz, E

    E. Schulz, E. M. Bender, M. Marelli, M. Botvinick, and S. J. Gershman. How should the advancement of large language models affect the practice of science?Proceedings of the National Academy of Sciences, 122(5):e2401227121, 2025

    2025

  39. [49]

    H. Seo, T. Hwang, J. Jung, H. Kang, H. Namgoong, Y . Lee, and S. Jung. Large language models as evaluators in education: Verification of feedback consistency and accuracy.Applied Sciences (2076-3417), 15(2), 2025

    2025

  40. [50]

    Shamsi, F

    K. Shamsi, F. Poursafaei, S. Huang, B. T. G. Ngo, B. Coskunuzer, and C. G. Akcora. Graphpulse: Topological representations for temporal graph property prediction. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  41. [51]

    S. Sia, D. Mueller, and K. Duh. Where does in-context learning happen in large language models? InNeurIPS, 2024

    2024

  42. [52]

    Singhal, S

    K. Singhal, S. Azizi, T.-J. Tu, S. Mahdavi, J. Wei, H. W. Chung, N. Scales, A. Tanwani, H. Cole, J. Lee, et al. Large language models encode clinical knowledge.Nature, 620(7973):172–180, 2023

    2023

  43. [53]

    Singhal, S

    K. Singhal, S. Azizi, T.-J. Tu, S. Mahdavi, J. Wei, H. W. Chung, N. Scales, A. Tanwani, H. Cole, J. Lee, et al. Towards expert-level medical question answering with large language models. arXiv preprint arXiv:2305.09617, 2023

    2023 arXiv

  44. [54]

    Souza, D

    A. Souza, D. Mesquita, S. Kaski, and V . Garg. Provably expressive temporal graph networks. Advances in neural information processing systems, 35:32257–32269, 2022

    2022

  45. [55]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin. Attention is all you need. InAdvances in Neural Information Processing Systems 30 (NeurIPS 2017), pages 5998–6008, 2017

    2017

  46. [56]

    J. Wang, K. Sun, L. Luo, W. Wei, Y . Hu, A. W. Liew, S. Pan, and B. Yin. Large language models-guided dynamic adaptation for temporal knowledge graph reasoning. In A. Globersons, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. M. Tomczak, and C. Zhang, editors,Advances in Neural...

    2024

  47. [57]

    W. Xia, M. Lai, C. Shan, Y . Zhang, X. Dai, X. Li, and D. Li. Explaining temporal graph models through an explorer-navigator framework. InThe Eleventh International Conference on Learning Representations, 2022

    2022

  48. [58]

    Y . Xia, D. Wang, Q. Liu, L. Wang, S. Wu, and X. Zhang. Chain-of-history reasoning for temporal knowledge graph forecasting. In L. Ku, A. Martins, and V . Srikumar, editors,Findings of the Association for Computational Linguistics, ACL 2024, Bangkok, Thailand and virtual meeti...

    2024

  49. [59]

    D. Xu, C. Ruan, E. Korpeoglu, S. Kumar, and K. Achan. Inductive representation learning on temporal graphs. InInternational Conference on Learning Representations (ICLR), 2020

    2020

  50. [60]

    A. Yang, B. Yang, B. Hui, B. Zheng, B. Yu, C. Zhou, C. Li, C. Li, D. Liu, F. Huang, G. Dong, H. Wei, H. Lin, J. Tang, J. Wang, J. Yang, J. Tu, J. Zhang, J. Ma, J. Xu, J. Zhou, J. Bai, J. He, J. Lin, K. Dang, K. Lu, K. Chen, K. Yang, M. Li, M. Xue, N. Ni, P. Zhang, P. Wang, R. ...

    2024 arXiv

  51. [61]

    A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Xia, X. Ren, X. Ren, Y ...

    2024 arXiv

  52. [62]

    M. Yang, M. Zhou, M. Kalander, Z. Huang, and I. King. Discrete-time temporal network embedding via implicit hierarchical learning in hyperbolic space. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1975–1985, 2021

    1975

  53. [63]

    R. Ye, C. Zhang, R. Wang, S. Xu, and Y . Zhang. Language is all a graph needs.arXiv preprint arXiv:2308.07134, 2023

    2023 arXiv

  54. [64]

    J. You, T. Du, and J. Leskovec. Roland: graph learning framework for dynamic graphs. In Proceedings of the 28th ACM SIGKDD conference on knowledge discovery and data mining, pages 2358–2366, 2022

    2022

  55. [65]

    L. Yu, L. Sun, B. Du, and W. Lv. Towards better dynamic graph learning: New architecture and unified library.Advances in Neural Information Processing Systems, 36:67686–67700, 2023

    2023

  56. [66]

    Zhang, Y

    X. Zhang, Y . Wang, X. Wang, and M. Zhang. Efficient neural common neighbor for temporal graph link prediction.arXiv preprint arXiv:2406.07926, 2024

    2024

  57. [67]

    Zhang, X

    Z. Zhang, X. Wang, Z. Zhang, H. Li, Y . Qin, and W. Zhu. Llm4dyg: can large language models solve spatial-temporal problems on dynamic graphs? InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 4350–4361, 2024

    2024

  58. [68]

    H. Zhao, Z. Wang, X. Wang, Y . Ma, Y . Lai, Z. Lin, Y . Zhang, and X. Cheng. Explainability for large language models: A survey.ACM Transactions on Intelligent Systems and Technology, 15(2):20:1–20:38, 2024

    2024

  59. [69]

    Zheng and A

    Z. Zheng and A. K. Bashir. Graph-enabled intelligent vehicular network data processing.IEEE Transactions on Intelligent Transportation Systems, 23(5):4726–4735, 2022. 14 A Limitations First, the process of converting temporal graphs into textual representations is inherently c...

    2022

  60. [1895]

    Association for Computational Linguistics, 2024

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.