Pith. sign in

REVIEW 3 major objections 6 minor 17 references

Towards Foundation Model on Temporal Knowledge Graph Reasoning

T0 review · 3 major / 6 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 This paper claims that a single pretrained model, POSTRA, can do zero-shot link prediction on temporal knowledge graphs whose entities, relations, and timestamps are all unseen, by transferring relative time-order and interaction patterns.

desk verdict A clean ULTRA-plus-temporal-encoding architecture whose central zero-shot claim is undermined by overlapping train/test datasets and test-time hyperparameter tuning. read the letter →

arxiv 2506.06367 v1 pith:6ZDNEEC7 submitted 2025-06-04 cs.AI

Jiaxin Pan , Mojtaba Nayyeri , Osama Mohammed , Daniel Hernandez , Rongchuan Zhang , Cheng Cheng , Steffen Staab This is my paper
classification cs.AI
keywords temporalknowledgegraphlinkpredictionfullyinductiveinferencezero-shottransferfoundationmodelsinusoidalpositionalencodingmessagepassingrelationinteraction
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

This paper proposes POSTRA, a model for temporal knowledge graph link prediction that works in a fully inductive setting: after training on one graph, it answers queries on graphs whose entities, relations, and timestamps never appeared in training. The authors argue that what transfers is not vocabulary-specific embeddings but relative structure, namely the order in which connected facts occur and the four universal ways relations can interact (head-to-head, head-to-tail, tail-to-head, tail-to-tail). They claim that a single pretrained POSTRA checkpoint improves zero-shot performance on multiple inductive temporal reasoning benchmarks, across daily, 15-minute, and yearly time granularities, and they support the transfer claim with a theorem showing that sinusoidal temporal embeddings preserve time-difference information independently of absolute timestamps. If correct, this is a step toward a foundation model for temporal knowledge graphs that would not need to be retrained for each new domain.

What carries the argument

The load-bearing machinery has three parts. First, a relation graph whose nodes are relations and whose edges are the four dataset-independent interaction types {h2h, h2t, t2h, t2t}; message passing over this graph produces transferable relation representations. Second, a sinusoidal positional encoding TE(i) for snapshot index i, which makes the Euclidean distance between two snapshots' embeddings depend only on the difference in their indices, not on the absolute times, so relative temporal order transfers across graphs with different calendars or granularities. Third, a quadruple encoder that runs two message-passing passes, one global over the whole graph and one local over a window around each query, then trades them off with a hyperparameter alpha and fuses the result with the temporal embedding through an MLP. The theoretical results show the temporal embedding is time-shift invariant and the affine scorer can represent multi-frequency periodic patterns.

What would settle it

Check the released cross-dataset splits for overlap: for instance, ICEWS14 covers the full year 2014 while ICEWS05-15 also contains 2014-01-01 through 2014-12-31, so the two graphs share timestamps unless explicit filtering was applied. If an audit finds any entity, relation, or timestamp present in both the training and inference graphs, or if removing the shared 2014 timestamps from the ICEWS05-15 test set erases most of POSTRA's advantage over ULTRA, the fully-inductive claim would be falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that temporal knowledge graph reasoning can be made fully inductive: entity, relation, and timestamp representations are generated on the fly from the query and the graph structure, so the same pretrained weights apply to a graph with a completely new vocabulary. POSTRA encodes each snapshot's position with sinusoidal functions, aggregates messages over relation-interaction graphs and entity graphs using local and global temporal contexts, and scores candidate answers with an MLP. In cross-dataset zero-shot experiments (training on one of ICEWS14, ICEWS05-15, GDELT, ICEWS18, or YAGO and testing on the others), POSTRA outperforms the prior inductive baselines INGRAM and ULTRA on average rank-based metrics, and it also beats LLM-based extrapolation models on ICEWS18 and YAGO while being cheap enough to run full ranking. The theoretical analysis proves that the temporal embedding makes distances depend only on relative time differences and that the scorer can express arbitrary periodic sequences with harmonically aligned frequencies.

Load-bearing premise

The load-bearing premise is that the training and inference graphs really are disjoint in entities, relations, and timestamps, so that any success on test graphs must come from transferred structure rather than from reusing elements seen during training; if even one timestamp or entity is shared across the split, the zero-shot framing is weakened.

Editorial extensions

If this is right

  • A single POSTRA checkpoint can be evaluated on a new temporal knowledge graph without retraining or fine-tuning, as long as the graph can be organized into snapshots.
  • The same model handles interpolation (transductive), extrapolation and forecasting (semi-inductive), and fully inductive settings, since it never relies on dataset-specific vocabularies.
  • Model size stays fixed as the number of entities, relations, or timestamps grows, because the only learned parameters are tied to the four relation-interaction types and the embedding dimension.
  • A model trained once on event data can be pointed at a new domain, such as encyclopedic facts, and yield usable link predictions without additional training.
  • Because the temporal representation is granularity-agnostic, the approach extends to graphs with mixed time units or irregular sampling, not just daily or yearly snapshots.

Reading between the lines

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

  • A testable extension: report the same cross-dataset results after explicitly removing every entity, relation, and timestamp shared between training and inference graphs; the paper describes disjointness but does not describe such filtering, and the raw date ranges of ICEWS14 and ICEWS05-15 overlap.
  • An implication the authors leave implicit: the relative-ordering mechanism could be applied to continuous event streams without fixed snapshots, by using sorted event indices or learned binning instead of calendar timestamps.
  • The foundation-model framing suggests a benchmark where disjointness is enforced by construction; until then, zero-shot numbers on overlapping benchmarks should be read as upper bounds on true inductive transfer.
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

3 major / 6 minor

Summary. The paper introduces POSTRA, a temporal knowledge graph embedding model intended for fully-inductive link prediction. The model uses sinusoidal positional encodings of timestamp indices, a relation-relation interaction graph with four universal interaction types, and global and local entity message passing to generate entity, relation, and temporal representations without dataset-specific vocabulary embeddings. The authors claim that a single pretrained POSTRA model transfers zero-shot across datasets with disjoint entities, relations, and timestamps, and they support this with cross-dataset experiments on ICEWS14, ICEWS05-15, GDELT, ICEWS18, and YAGO, plus theoretical analyses of time-shift invariance and periodic pattern capture. The central claim is that this is the first fully-inductive temporal knowledge graph reasoning approach and a step toward a foundation model for temporal KGs.

Significance. If the fully-inductive zero-shot claim were established, the contribution would be significant: it would move temporal KG reasoning beyond transductive and semi-inductive settings, enable cross-dataset transfer without retraining, and address temporal granularity and time-span discrepancies. The model design is clean, parameter-efficient, and the basic idea of using sinusoidal temporal encodings with message passing is sensible. The paper also contains machine-checkable mathematical statements, and Theorem 1 is correctly proved. However, the empirical foundation is currently undermined by dataset overlap, and one of the central theorems has a proof gap. These issues must be resolved before the claims can be accepted as evidence for a foundation-model-style transfer.

major comments (3)
  1. [Section 5, Table 6] The assertion that cross-dataset evaluations ensure 'no overlap in entities, relations, or timestamps' is contradicted by the paper's own statistics. ICEWS14 has 365 daily timestamps covering 2014, while ICEWS05-15 has 4017 daily timestamps covering 2005–2015, so Ttrain ∩ Tinf contains all 365 ICEWS14 days; GDELT (Apr 2015–Mar 2016, 366 timestamps) overlaps ICEWS05-15 in 2015. ICEWS14, ICEWS05-15, and ICEWS18 are drawn from the same ICEWS event ontology, so the entity and relation vocabularies are likely shared (e.g., 230 versus 251 relations). No filtering procedure is described, and no split files are provided. Because POSTRA conditions representations on the relation graph and propagates over entity graphs, any shared relation, entity, or timestamp allows the model to replay patterns learned from Gtrain; test quadruples from ICEWS05-15 that also appear in ICEWS14 training data could even be scored by memorization. The zero-shot gains in Table 1 therefore do not support fully-inductive transfer as defined in Section 2 without additional evidence of genuine disjointness.
  2. [Appendix I, Theorem 2] The proof of C⇒S is incomplete. Step 2 derives w^T (TE(τ+P_i) − TE(τ)) = 0 only for τ ∈ {0,...,T−1}; the text then says 'replacing τ by τ+P_i and iterating k times' establishes periodicity on all integers, but for arbitrary real frequencies and P_i > 1 the replacement requires rows of B(i) that are not in the system. For example, with T=4, P_i=2, the equations imply w^T TE(4)=w^T TE(2)=w^T TE(0) and w^T TE(5)=w^T TE(3)=w^T TE(1), but they impose no relation between w^T TE(6) and w^T TE(4); the claimed all-integer periodicity does not follow. Since Theorem 2 is the stated theoretical basis for the claim that the scorer can capture and express multiple periodicities even with a linear f_θ, either the theorem must be proved under the stated generality, or the claim should be explicitly restricted to the harmonically aligned frequencies treated in Theorem 3.
  3. [Section 2, Task Formulation] The fully-inductive evaluation protocol is weakened by the statement that at inference the model uses QO to compute embeddings and 'tune hyperparameters based on the model's performance on Qvalid' (Section 2). This means target-graph validation data are used for model selection (e.g., α and k are dataset-specific in Appendix F). This is not a fully zero-shot procedure in the usual sense, and it should be reconciled with Section 5's claim that 'no fine-tuning is applied during testing.' If the intended meaning is only that learned weights are frozen while hyperparameters are chosen per target graph, that should be stated precisely, and the hyperparameter values used in each Table 1 cell should be reported.
minor comments (6)
  1. [Section 1, Figure 1 caption] The caption contains a typo: 'Fully-Inductive Interence' should be 'Fully-Inductive Inference.'
  2. [Section 4] In Section 4.1, the sentence describing relation message passing contains the typo 'meassage-passing'; it should read 'message-passing.'
  3. [Section 2] The notation 'Vtrain ∩Vinf =/ 0' is ambiguous; the authors likely intend the empty set symbol ∅. As written, it is inconsistent with the surrounding text claiming complete disjointness, so it should be corrected.
  4. [Table 3 vs Table 1] The ablation 'w/o TE' in Table 3 reports MRR 11.1, H@1 12.0, H@10 15.0, which exactly matches the ULTRA row for training on ICEWS14 and testing on ICEWS05-15 in Table 1. The authors should clarify whether this is a coincidence or whether removing the temporal encoding reduces POSTRA to the ULTRA baseline; if the latter, this should be stated explicitly.
  5. [Section 6.1, Observation 1] The text states that the time span of the experimental datasets 'ranges from just 1 month (GDELT) to 189 years (YAGO),' but Section 5 and Table 6 describe GDELT as covering one year (April 2015–March 2016) with 366 timestamps. This inconsistency should be corrected.
  6. [Reproducibility] The paper mentions in Section 7 that data splits are provided in 'our codebase,' but no code or data link is given. Given that the fully-inductive claim hinges on the exact composition of training and inference graphs, the split files and filtering procedures (if any) must be released or described in detail.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the theorems are mathematical properties of the chosen encodings and representational equivalences, and the ULTRA backbone is external prior work.

full rationale

I examined the claimed derivation chain. The temporal-transferability guarantee rests on Theorem 1, which is a trigonometric identity of the sinusoidal embedding the authors explicitly chose; it is not fitted from data and does not by itself produce the reported zero-shot scores. Theorems 2 and 3 are representation-capacity statements: they show equivalence between a linear-system compatibility condition (C) and the existence of an affine scorer with specified periodicity (S), and give a sufficient frequency choice for solvability. These are internal mathematical results about the model family, not empirical predictions reduced to training outputs. The relation-encoding component is adopted, with citation, from ULTRA (Galkin et al., 2024), an external prior method, and the sinusoidal encoding from Transformer literature (Vaswani et al., 2017); neither is a self-citation. The only author self-citation (HGE, Pan et al., 2024) appears in the related-work survey and is not load-bearing for any claim. The skeptical concern about overlapping entities/relations/timestamps across the cross-dataset splits (ICEWS14 lies inside ICEWS05-15; GDELT overlaps ICEWS05-15) is a serious threat to the validity of the zero-shot evaluation, but it is not a circularity of the derivation: no equation in the paper makes a test score equal to a training fit by construction, and no fitted parameter is renamed as a prediction. Accordingly, no circular step meets the required evidence bar.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on three domain assumptions: (1) the four ULTRA relation interactions are universal and transferable, (2) relative snapshot order (not absolute time) is the right transferable temporal signal, and (3) NBFNet-style message passing generalizes across graphs. The paper also assumes the experiments implement disjoint vocabularies, which they do not. No new entities or forces are introduced.

free parameters (3)
  • alpha (local vs global weight) = 0.5 (ICEWS14), 0.8 (ICEWS0515, GDELT)
    Equation 6; tuned on the target validation set at inference, which weakens the zero-shot claim.
  • k (local time window size) = 0 (ICEWS14), 1 (ICEWS0515, GDELT)
    Section 4.3 and Table 8; controls the local graph span and is dataset-dependent.
  • beta (sinusoidal frequency base) = 10000
    Section 4.2 and Table 8; fixed across datasets after sensitivity analysis showed robustness.
assumptions (4)
  • domain assumption The four relation interaction types (h2h, h2t, t2h, t2t) are universal and transferable across TKGs.
    Section 4.1, adopted from ULTRA (Galkin et al. 2024); if target relations do not follow these interactions, the relation representation transfer fails.
  • domain assumption Relative temporal ordering between snapshots is the transferable temporal signal, so positional encoding over snapshot indices suffices.
    Section 4.2; the model discards absolute time values and granularity, an explicit design choice.
  • domain assumption NBFNet-style message passing on local and global entity graphs yields representations that transfer to graphs with entirely new vocabularies.
    Section 4.3; the inductive capability of the GNN is assumed rather than proved.
  • ad hoc to paper The experimental train/test pairs have disjoint entities, relations, and timestamps as defined in Section 2.
    Stated in Section 5 ('no overlap in entities, relations, or timestamps') but contradicted by Table 6: ICEWS14 is a subset of the ICEWS05-15 time range.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Foundation Model on Temporal Knowledge Graph Reasoning." pith.science (2026). https://pith.science/paper/6ZDNEEC7

@misc{pith2026250606367,
  author       = {Pith},
  title        = {Pith review of: Towards Foundation Model on Temporal Knowledge Graph Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZDNEEC7}},
  note         = {Machine review of arXiv:2506.06367}
}
read the original abstract

Temporal Knowledge Graphs (TKGs) store temporal facts with quadruple formats (s, p, o, t). Existing Temporal Knowledge Graph Embedding (TKGE) models perform link prediction tasks in transductive or semi-inductive settings, which means the entities, relations, and temporal information in the test graph are fully or partially observed during training. Such reliance on seen elements during inference limits the models' ability to transfer to new domains and generalize to real-world scenarios. A central limitation is the difficulty in learning representations for entities, relations, and timestamps that are transferable and not tied to dataset-specific vocabularies. To overcome these limitations, we introduce the first fully-inductive approach to temporal knowledge graph link prediction. Our model employs sinusoidal positional encodings to capture fine-grained temporal patterns and generates adaptive entity and relation representations using message passing conditioned on both local and global temporal contexts. Our model design is agnostic to temporal granularity and time span, effectively addressing temporal discrepancies across TKGs and facilitating time-aware structural information transfer. As a pretrained, scalable, and transferable model, POSTRA demonstrates strong zero-shot performance on unseen temporal knowledge graphs, effectively generalizing to novel entities, relations, and timestamps. Extensive theoretical analysis and empirical results show that a single pretrained model can improve zero-shot performance on various inductive temporal reasoning scenarios, marking a significant step toward a foundation model for temporal KGs.

Figures

Figures reproduced from arXiv: 2506.06367 by the authors.

Figure 1
Figure 1. (b) depicts the semi-inductive setting (temporal knowledge graph extrapolation or forecasting), where the timestamp information in the test graph is not available during training. However, the entities and/or relations remain consistent between the training and test graphs. China Barack Obama Mileva Maric Marries 1903 Lives in 1905 Albert Einstein Training Graph Seen entity, relation and time Unseen entity, relation… view at source ↗
Figure 2
Figure 2. The bottom TKGs depict temporal knowledge graphs from different domains. The upper relation graphs show relative relation rep￾resentations learned via fundamental interaction patterns (See Section 4.1) and the relative tempo￾ral ordering between corresponding facts (∆t) (See Section 4.2) which are transferrable across TKGs. More details are shown in Section E in Appendix. Let V be a finite set of entities, R be a fi… view at source ↗
Figure 3
Figure 3. The overall architecture of POSTRA. The model first constructs a Global Relation Graph [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Relative Representation Transfer F Hyperparameter and Training Details Our experiments were conducted on 4 NVIDIA A100 GPUs with 48GB of RAM. We set the maximal training epoch as 10 and negative samples as 512. We use a batch size of 16, 2, and 1 for training ICEWS14, …
Figure 5
Figure 5. Figure 5: MRR and H@10 performance with different k, α and β. We report ICEWS14’s pre-training performance and keep all other hyper-parameters as the fixed best setting. H Case Study To intuitively demonstrate POSTRA’s capability in generating compact and distinguishable embed￾d…
Figure 6
Figure 6. Figure 6: PCA visualizations of five quadruples with (China, Host a visit, ?, ?) [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

  1. [3]

    Learning sequence encoders for temporal knowledge graph completion

    Alberto García-Durán, Sebastijan Dumanˇci´c, and Mathias Niepert. Learning sequence encoders for temporal knowledge graph completion. In Ellen Riloff, David Chiang, Julia Hockenmaier, and Jun’ichi Tsujii, editors,Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 4816–4821, Brussels, Belgium, October-November

    work page 2018

  2. [6]

    In contrast, the Local Quadruple Representation is more sensitive to immediate temporal context, resulting in more distinctive embeddings across different time points

    The visualization reveals that the Global Quadruple Representation tends to capture general relational patterns, producing embeddings that are closely clustered for the same query structure (China, Host a Visit, ?, ?). In contrast, the Local Quadruple Representation is more sensitive to immediate temporal context, resulting in more distinctive embeddings ...

    work page 2014

  3. [8]

    Temporal knowledge graph forecasting without knowledge using in-context learning

    Dong-Ho Lee, Kian Ahrabian, Woojeong Jin, Fred Morstatter, and Jay Pujara. Temporal knowledge graph forecasting without knowledge using in-context learning. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 544–557, 2023a. Jaejun Lee, Chanyoung Chung, and Joyce Jiyoung Whang. Ingram: Inductive knowledge graph ...

    work page 2023

  4. [9]

    Indigo: Gnn-based inductive knowledge graph completion using pair-wise encoding.Advances in Neural Information Processing Systems, 34:2034–2045,

    Shuwen Liu, Bernardo Grau, Ian Horrocks, and Egor Kostylev. Indigo: Gnn-based inductive knowledge graph completion using pair-wise encoding.Advances in Neural Information Processing Systems, 34:2034–2045,

    work page

  5. [11]

    Timetraveler: Reinforcement learning for temporal knowledge graph forecasting

    12 Haohai Sun, Jialun Zhong, Yunpu Ma, Zhen Han, and Kun He. Timetraveler: Reinforcement learning for temporal knowledge graph forecasting. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8306–8319,

    work page 2021

  6. [12]

    Riemanngfm: Learning a graph foundation model from structural geometry

    Li Sun, Zhenhao Huang, Suyang Zhou, Qiqi Wan, Hao Peng, and Philip S Yu. Riemanngfm: Learning a graph foundation model from structural geometry. InTHE WEB CONFERENCE 2025,

    work page 2025

  7. [13]

    Pre-trained language model with prompts for temporal knowledge graph completion

    Wenjie Xu, Ben Liu, Miao Peng, Xu Jia, and Min Peng. Pre-trained language model with prompts for temporal knowledge graph completion. InFindings of the Association for Computational Linguistics: ACL 2023, pages 7790–7803,

    work page 2023

  8. [14]

    Embedding entities and relations for learning and inference in knowledge bases

    Bishan Yang, Scott Wen-tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. Embedding entities and relations for learning and inference in knowledge bases. InProceedings of the International Conference on Learning Representations (ICLR) 2015,

    work page 2015

Show all 17 references
  1. [15]

    13 A Parameter Count and Complexity From Table 5, we observe that the number of parameters in POSTRA is agnostic to the dataset size. Since POSTRA does not initialize embeddings based on |V|, |R|, or |T| , but instead relies on a fixed number of relation interactions |H|, its ...

    2021

  2. [16]

    15 Table 8: Detailed hyperparameters

    For the baseline methods, we utilized the settings in their original papers. 15 Table 8: Detailed hyperparameters. GNNr denotes the relation representation encoder, GNNq is quadruple representation encoder. Module Hyperparameter Pre-training GNNr # layers 6 hidden dim 64 MSG D...

  3. [2015]

    Deriving validity time in knowledge graph

    Julien Leblay and Melisachew Wudage Chekol. Deriving validity time in knowledge graph. In Companion proceedings of the the web conference 2018, pages 1771–1776,

    2018

  4. [2018]

    doi: 10.18653/v1/D18-1516

    Association for Computational Linguistics. doi: 10.18653/v1/D18-1516. URL https://aclanthology.org/ D18-1516/. Julia Gastinger, Timo Sztyler, Lokesh Sharma, Anett Schuelke, and Heiner Stuckenschmidt. Compar- ing apples and oranges? on the evaluation of methods for temporal kno...

    doi

  5. [2020]

    Ecola: Enhancing temporal knowledge embeddings with contextualized language representations

    11 Zhen Han, Ruotong Liao, Jindong Gu, Yao Zhang, Zifeng Ding, Yujia Gu, Heinz Koeppl, Hinrich Schütze, and V olker Tresp. Ecola: Enhancing temporal knowledge embeddings with contextualized language representations. InFindings of the Association for Computational Linguistics: ...

    2023

  6. [2021]

    doi: 10.18653/v1/2021.acl-long.520

    Association for Computational Linguistics. doi: 10.18653/v1/2021.acl-long.520. URLhttps://aclanthology.org/2021.acl-long.520/. Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding.Neurocomputin...

    2021 doi

  7. [2023]

    Recurrent event network: Autoregressive structure inferenceover temporal knowledge graphs

    Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. Recurrent event network: Autoregressive structure inferenceover temporal knowledge graphs. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6669–6683,

    2020

  8. [2024]

    Kai Chen, Ye Wang, Yitong Li, Aiping Li, Han Yu, and Xin Song

    URL https://arxiv.org/ abs/2403.04782. Kai Chen, Ye Wang, Yitong Li, Aiping Li, Han Yu, and Xin Song. A unified temporal knowledge graph reasoning model towards interpolation and extrapolation. InProceedings of the 62nd Annual Meeting of the Association for Computational Lingu...

    arXiv

  9. [2025]

    zrllm: Zero-shot relational learning on temporal knowledge graphs with large language models

    Zifeng Ding, Heling Cai, Jingpei Wu, Yunpu Ma, Ruotong Liao, Bo Xiong, and V olker Tresp. zrllm: Zero-shot relational learning on temporal knowledge graphs with large language models. InProceedings of the 2024 Conference of the North American Chapter of the Association for Com...

    2024

Pith tools

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