Pith. sign in

REVIEW 4 major objections 5 minor 30 references

RECIPE-TKG: From Sparse History to Structured Reasoning for LLM-based Temporal Knowledge Graph Completion

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read The paper argues that LLM-based temporal knowledge graph completion is bottlenecked not by model capacity but by shallow history retrieval and evaluation choices, and that a three-part framework — rule-based multi-hop sampling…

desk verdict Useful framework, but the filter's credit is probably a historical fallback in disguise, and the LLM baseline comparison is incomplete. read the letter →

arxiv 2505.17794 v1 pith:JJ4VEJT3 submitted 2025-05-23 cs.LG

classification cs.LG
keywords temporalknowledgegraphcompletionlargelanguagemodelssparsehistorymulti-hopretrievalcontrastivefine-tuningtest-timefilteringLoRAadapterslinkforecasting
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 claims that LLM-based temporal knowledge graph completion fails most when a query's true answer is absent from the one-hop history the prompt provides, and that this failure is fixable without retraining the base model. To that end it introduces RECIPE-TKG, a three-part framework: a rule-based two-stage sampler that pulls in multi-hop, temporally weighted historical facts; contrastive fine-tuning of small LoRA adapters that pushes embeddings of relationally compatible entities together; and a test-time filter that measures cosine similarity between a generated answer and the prompt context, regenerating or falling back to history-grounded candidates when similarity is low. On ICEWS14, ICEWS18, GDELT, and YAGO the method reports relative Hits@1/3/10 gains of 8% to 30.6% over prior LLM baselines, with the largest margins in sparse-history regimes and new state-of-the-art results on ICEWS14 and YAGO. The paper also argues, via a controlled re-evaluation, that much of the apparent advantage of a previous fine-tuned method comes from its evaluation pipeline and sampling choices rather than from fine-tuning itself.

What carries the argument

The argument is carried by a three-component pipeline rather than a single identity. The sampler's load-bearing object is the composite weight $w = w_n \cdot w_f \cdot (w_t + w_c + w_{cp})$ over candidate quadruples, which multiplicatively combines reachability, anti-frequency, temporal recency, subject/relation co-occurrence, and connectivity to the initial rule-based context, so structurally distant but relevant facts enter the prompt. The learner uses a margin-based contrastive loss $\mathcal{L}_{\text{contrastive}} = \frac{1}{N_c} \sum_{i=1}^{N_c} \max(0, \|a_i - \text{pos}_i\|_2 - \|a_i - \text{neg}_i\|_2 + m)$ over relation-labeled positive/negative pairs, computed on attention-aggregated entity embeddings. The refiner applies the decision rule $p' = p$ if $p \in H$ or $\phi(p,c) \ge \tau$, otherwise regenerate, otherwise $\arg\max_{h \in H} \psi(h)$ with $\psi(h) = \beta f(h) + (1-\beta) r(h)$, where $\phi(p,c)$ is the cosine similarity from the frozen all-mpnet-base-v2 sentence transformer. The ablation attributes the single largest Hits@10 jump (0.580 to 0.651 on ICEWS14) to this final filtering step.

What would settle it

On each dataset, re-estimate the filtering threshold $\tau$ on the test predictions themselves and on a scrambled (prediction, context) pairing; if the Hits@10 gain from filtering (0.580 to 0.651 on ICEWS14) shrinks to noise once $\tau$ is re-tuned on the test distribution, the filter is exploiting threshold over-fitting rather than semantic alignment. The paper reports the relevant similarity gap only for ICEWS14 ($\Delta\mu = 0.057$ between correct and incorrect predictions), so repeating the same histogram analysis on GDELT and YAGO — where filtering is credited with a 16.8% Hits@10 gain — would settle whether the assumption transfers.

Watch

Extended reading notes

Core claim

The central discovery is that the apparent gains of LLM-based TKG forecasting are dominated by how history is retrieved and how outputs are post-processed, not by how well the model is trained: under a unified evaluation harness, switching the evaluation pipeline alone raises Hits@1 from 25.8% to 34.4%, while switching sampling adds roughly one more point and fine-tuning adds about 1.3 points. RECIPE-TKG turns this diagnosis into a method. First, rule-based multi-hop history sampling augments subject-anchored temporal-rule facts with a weighted multi-hop expansion that favors reachable, recent, low-frequency, and structurally connected quadruples. Second, contrastive fine-tuning with LoRA adapters, using relation-guided positive and negative neighbor pairs and a margin loss on attention-aggregated entity embeddings, sharpens the latent space so semantically compatible entities cluster. Third, test-time semantic filtering accepts a generation only if its sentence-embedding cosine similarity to the prompt exceeds a tuned threshold or the entity already appears in retrieved history; otherwise it regenerates and ultimately falls back to a frequency-and-recency-scored historical candidate. The paper reports that this combination yields relative improvements of 8.0% to 30.6% over the best LLM baselines across the four benchmarks, lifts Hits@10 above 60% even in the sparsest history bins where prior methods stay below 30%, and produces semantically coherent near-miss predictions even when exact matches are absent from history.

Load-bearing premise

The largest single accuracy jump comes from the test-time filter, and that filter assumes that the cosine similarity between a generated answer and the prompt text — computed with one fixed sentence transformer — reliably separates good from bad predictions, and that the single threshold tuned on a development set keeps working on test queries without retuning.

Editorial extensions

If this is right

  • On ICEWS14 and YAGO the method reports new state-of-the-art scores among all compared methods — up to 11.9% relative improvement over the strongest competitor — and on ICEWS18 it beats the best LLM baseline by 30.6% relative Hits@10 while matching the top embedding-based method.
  • Sparse-history queries are not inherently harder: with only 0–2 retrieved facts, Hits@10 rises above 60% where previous LLM methods sit below 30%, so retrieval depth, not model capacity, was the binding constraint.
  • Because the controlled re-evaluation shows that most of GenTKG's reported gain over ICL comes from evaluation postprocessing and sampling rather than supervised fine-tuning, future LLM-based TKG comparisons need a standardized decoding-and-evaluation harness before attributing gains to reasoning.
  • Contrastive tuning produces semantically plausible near-miss predictions even when the exact gold entity is absent from history, which improves ranked candidate lists in forecasting and decision-support settings.
  • The framework needs only a few thousand training shots and LoRA adapters, and it transfers to a newer base model (LLaMA-3-8B) with comparable results, supporting deployment without full model retraining.

Reading between the lines

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

  • The method redirects compute from training to retrieval and verification, suggesting a template for other knowledge-intensive generation tasks: keep the base model frozen, enrich the prompt with graph-structured evidence, and gate outputs by embedding similarity to the context.
  • Because cosine similarity to the prompt is treated as a correctness proxy, the filter amounts to a generic hallucination guard; a natural transfer test is to apply the same thresholded filtering to closed-book question answering or fact checking, where context-misaligned generations are also the dominant error type.
  • The re-evaluation result (evaluation pipeline moves Hits@1 by +8.6 points, fine-tuning by +1.3) implies that published LLM-based TKG comparisons may be comparing pipelines rather than models; a shared evaluation harness, not a new model, would be the cheapest next contribution to the field.
  • The paper's own limitations — clean, fully observed graphs and rule mining that must be rerun whenever the TKG changes — point to the stress test that would bound the approach's usefulness: injecting dropped or delayed events into the history and measuring how the sampling weights and the filter threshold degrade.
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 / 5 minor

Summary. The paper proposes RECIPE-TKG, a three-component framework for LLM-based temporal knowledge graph completion: rule-based multi-hop history sampling (RBMH), contrastive fine-tuning of LoRA adapters with a relation-guided contrastive loss, and test-time semantic filtering with a historical fallback. The method is evaluated on ICEWS14, ICEWS18, GDELT, and YAGO using LLaMA2-7B (and LLaMA3-8B in a case study) against embedding-based, rule-based, and two LLM-based baselines (ICL and GenTKG). The authors report consistent improvements on most metrics, with the largest gains in Hits@10 and in low-history settings, and they also re-evaluate prior baselines under a unified setup to show that much of GenTKG's reported gain comes from evaluation and sampling choices.

Significance. If the central claims hold, RECIPE-TKG is a useful modular contribution for LLM-based TKG forecasting in sparse-history conditions, and the re-evaluation of prior baselines under consistent evaluation is valuable. The paper is also commendable for releasing code and for providing hyperparameter-sensitivity analyses. However, the significance is currently limited by three weaknesses: the abstract overstates the GDELT results, the largest single contribution (test-time filtering) is not isolated from the historical fallback, and recent LLM-based baselines are omitted from the comparison. These gaps, if addressed, would substantially strengthen the paper.

major comments (4)
  1. [Abstract and Section 5.2, Table 2] The abstract and introduction claim that RECIPE-TKG 'outperforms previous LLM-based approaches' with 'relative gains on Hits@1/3/10 ranging from 8% to 30.6%.' Table 2 shows this is not accurate for GDELT: RECIPE-TKG's Hits@1 is 0.095 versus 0.099 for GenTKG (a relative decrease of 4.0%) and Hits@3 is 0.192 versus 0.193 (relative decrease of 0.5%); only Hits@10 improves. The headline claim should be qualified to reflect that the method is better on most metrics but not uniformly so.
  2. [Section 4.3, Appendix C, Table 4] The test-time filtering contribution is not isolated from the historical fallback. In Eq. 8, a prediction is accepted if it appears in H or if cosine similarity exceeds tau; otherwise the model regenerates once (k=1) and then falls back to a frequency/recency scorer psi(h) in Eq. 9. Appendix C reports that only 1.5% of non-historical predictions are correct and that 9.1% of predictions are non-historical while the gold entity is in H. These facts imply that replacing low-similarity predictions with the historical fallback alone would correct most of these errors. Indeed, Table 4 shows the filter raises Hits@10 from 0.580 to 0.651 while Hits@1 moves only 0.392 to 0.393, a pattern consistent with the fallback filling lower-ranked slots. To substantiate the claim that the semantic similarity threshold is load-bearing, the authors must (i) report per-dataset tau and beta values and the development-set sizes, and (ii) ablate the similarity test against a random acceptance rule matched on acceptance rate, or against a rule that always accepts predictions in H and falls back after one regeneration. Without this, the paper does not establish that the embedding-based filter contributes beyond the hard 'p in H' acceptance and the historical fallback.
  3. [Section 1 and Section 5.1 (Baselines), Table 2] The paper cites Luo et al. (2024) and Xia et al. (2024) as recent LLM-based TKG methods in the introduction, but Table 2 compares only ICL (Lee et al., 2023) and GenTKG (Liao et al., 2024). The claim of state-of-the-art among LLM-based methods on ICEWS14 and YAGO is therefore only with respect to two baselines. The authors should include these recent methods (or explicitly justify their exclusion), since the reader cannot otherwise evaluate the competitiveness of the proposed approach.
  4. [Section 5.1 and Table 2] The main results are reported without error bars or multiple runs. Given the near-ties on GDELT (e.g., Hits@1 0.095 vs 0.099) and the stochastic components (LoRA training, sampling, generation), the reader cannot assess whether the negative deltas are significant. Reporting mean and standard deviation over at least three seeds, or providing significance tests, is necessary to support the claimed improvements.
minor comments (5)
  1. [Section 6.1, Table 4 paragraph] The sentence 'combining test-time filtering with RBMH Sampling and Test-time Filtering (RECIPE-TKG)' appears to have a typo; it should likely read 'combining RBMH Sampling, contrastive fine-tuning, and test-time filtering (RECIPE-TKG).'
  2. [Table 1 caption] The caption does not specify the dataset on which the re-evaluation was performed; the text implies ICEWS14 but this should be stated explicitly.
  3. [Section 6.1] The text says 'in-context learning (ICT)' but the abbreviation should be ICL, as used elsewhere in the paper.
  4. [Appendix C] The description of k is ambiguous: the text says 'we use k=1' but Eq. 8 says 'after k attempts.' Please clarify whether k denotes the number of regeneration attempts after the initial generation or the total number of attempts.
  5. [Appendix B.4] There is a missing space in 'fine-tuneLLaMA-2-7B'; it should read 'fine-tune LLaMA-2-7B.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework's components are tested against external baselines and independent test labels, and no prediction reduces to its inputs by construction.

full rationale

RECIPE-TKG's three components are not defined in terms of the target metrics. Rule-based multi-hop sampling adapts the external TLR algorithm and adds explicitly defined weighting components; contrastive fine-tuning uses relation labels from an external GPT-4o system and a standard contrastive loss; test-time filtering uses a frozen external sentence transformer (all-mpnet-base-v2), with the threshold tau optimized on a development set rather than on test labels, and the fallback is a transparent frequency/recency score over retrieved history. The headline results compare against independent embedding-based, rule-based, and LLM-based baselines using held-out test splits. The only self-citation (Wang et al., 2025, on which one co-author appears) is used in a general statement about LLM reasoning benchmarks and is not load-bearing. The large Hits@10 gain attributed to filtering may partly reflect the historical fallback mechanism rather than semantic similarity, but that is a potential experimental confound, not a circular reduction: the filter's output is not equivalent to its input by definition, and the evaluation remains grounded in external test labels. Therefore no circular step is established, and the derivation chain is self-contained with respect to the concerns enumerated in the circularity analysis.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a set of hand-chosen hyperparameters (thresholds, loss weights, sampling weights), domain assumptions about data cleanliness and the reliability of GPT-4o relation labels and sentence-transformer similarity, and a heuristic relational prior. No machine-checked proof or shipped reproduction artifacts back the empirical claims.

free parameters (7)
  • tau (semantic similarity threshold) = learned on dev set; ~0.6 for ICEWS14 (Figure 11)
    Threshold in test-time filtering that accepts or rejects a generated prediction; direct contributor to the filtering gain in Table 4.
  • alpha (contrastive loss weight) = 0.2
    Balances contrastive loss and cross-entropy loss in Eq. 4; sensitivity shown in Figure 10.
  • margin m = 1.0
    Margin in contrastive loss (Eq. 5) that separates positive and negative pairs.
  • beta (fallback balance) = 0.6
    Blends frequency and recency scoring in the historical fallback (Eq. 9).
  • gamma_1 to gamma_4 (sampling weights) = 0.6, 0.6, 0.01, 0.1
    Control hop decay, frequency suppression, time decay, and co-occurrence smoothing in the multi-hop sampling weight (Eq. 2).
  • N (max history length) = 50
    Maximum facts fed to the LLM, set by context window.
  • regeneration attempts k = 1
    Number of regeneration tries before falling back to historical scoring.
assumptions (4)
  • domain assumption Temporal knowledge graphs are clean and fully observed; all historical events are available and correct.
    Stated in Limitations: the framework is 'built on clean, fully observed temporal knowledge graphs'.
  • domain assumption GPT-4o relation classification into positive/negative/neutral is accurate enough for contrastive pair construction.
    Section 4.2 and Appendix B.1 rely on this classification, but no accuracy or consistency metrics are reported.
  • domain assumption Sentence-transformer cosine similarity between prediction and prompt is a reliable proxy for correctness.
    Motivated by Figure 4 and Appendix C, where the mean similarity difference between correct and incorrect predictions is only 0.057 on ICEWS14; no per-dataset validation.
  • ad hoc to paper The 'enemy of my enemy is my friend' heuristic is a valid relational prior for geopolitical TKGs.
    Introduced in Section 4.2 to justify contrastive pairs; not derived from data or theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RECIPE-TKG: From Sparse History to Structured Reasoning for LLM-based Temporal Knowledge Graph Completion." pith.science (2026). https://pith.science/paper/JJ4VEJT3

@misc{pith2026250517794,
  author       = {Pith},
  title        = {Pith review of: RECIPE-TKG: From Sparse History to Structured Reasoning for LLM-based Temporal Knowledge Graph Completion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JJ4VEJT3}},
  note         = {Machine review of arXiv:2505.17794}
}
read the original abstract

Temporal Knowledge Graphs (TKGs) represent dynamic facts as timestamped relations between entities. TKG completion involves forecasting missing or future links, requiring models to reason over time-evolving structure. While LLMs show promise for this task, existing approaches often overemphasize supervised fine-tuning and struggle particularly when historical evidence is limited or missing. We introduce RECIPE-TKG, a lightweight and data-efficient framework designed to improve accuracy and generalization in settings with sparse historical context. It combines (1) rule-based multi-hop retrieval for structurally diverse history, (2) contrastive fine-tuning of lightweight adapters to encode relational semantics, and (3) test-time semantic filtering to iteratively refine generations based on embedding similarity. Experiments on four TKG benchmarks show that RECIPE-TKG outperforms previous LLM-based approaches, achieving up to 30.6\% relative improvement in Hits@10. Moreover, our proposed framework produces more semantically coherent predictions, even for the samples with limited historical context.

Figures

Figures reproduced from arXiv: 2505.17794 by the authors.

Figure 1
Figure 1. Example of LLM-based TKG reasoning. Prior [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Prediction failures under sparse or shallow history. (a) Accuracy vs. history length shows longer contexts support better reasoning. (b) Most non-historical targets require multi-hop reasoning, but are unreachable with 1-hop sampling. (c) Accuracy drops sharply on non-historical predictions for both ICL and SFT. 2.2 Limitations of Supervised Fine-Tuning Supervised fine-tuning (SFT) is widely used to adapt LLMs to TK… view at source ↗
Figure 3
Figure 3. Overview of RECIPE-TKG. RECIPE-TKG follows a three-stage framework: (1) History Sampling, which retrieves query-relevant facts via a two-phase strategy combining rule-based retrieval and context-guided expansion; (2) Contrastive Learning, which jointly optimizes entity embeddings using contrastive and cross￾entropy losses. Positive/negative pairs are sampled from the subgraph, and embeddings are generated via a lear… view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Distribution of semantic similarity values for [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Hits@10 grouped by number of historical facts. RECIPE-TKG consistently outperforms ICL and GenTKG across all history lengths, with particularly strong improvements when the input history is sparse. historical context is sparse or entirely absent. As discussed in Append…
Figure 6
Figure 6. Figure 6: Performance of ICL-RBMH under different sampling hyperparameter configurations. where ETLR is the set of all 1-hop neighbors identi￾fied in the TLR stage. This guides the expansion toward semantically coherent subgraphs. A.2.3 Hyperparameter Sensitivity Experiment [PI…
Figure 7
Figure 7. Figure 7: Prompt used for relation classification. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Distribution of relation types in four datasets [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Instruction-style prompt format for TKG forecasting. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Effect of contrastive weight (α) produces final entity embeddings, jointly trained with the model. B.5 Hyperparameter Sensitivity Experiment As shown in [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: Effect of filtering threshold (τ ) correct (C) and incorrect (I) prediction similarities: τ ∗ = arg max τ [FC(τ ) − FI(τ )] (11) where F is the empirical CDF of cosine similarity values over samples from C and I. Fallback Scoring. If generation fails after k it￾eratio…
Figure 12
Figure 12. Figure 12: Semantic distance (ϕ) vs. history length on ICEWS14 under TLR sampling. The encircled region highlights CL’s improved semantic grounding in sparse-history settings [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Semantic distance (ϕ) vs. history length for the same TLR-identified sparse samples, but evaluated under RBMH Sampling. The model exhibits more stable behavior across history lengths [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Semantic distance (ϕ) vs. history length for the same TLR-identified sparse samples, but evaluated under RBMH Sampling. CFT learns better with RBMH as it samples the deeper relationships between entities. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Top-10 predictions from four models. RECIPE-TKG produce semantically closer outputs to the ground [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 13 canonical work pages

  1. [1]

    Elizabeth Boschee, Jennifer Lautenschlager, Sean O’Brien, Steve Shellman, James Starz, and Michael Ward. 2015. https://doi.org/10.7910/DVN/28075 ICEWS Coded Event Data

  2. [2]

    Rochana Chaturvedi. 2024. https://dl.acm.org/doi/10.1145/3589335.3651256 Temporal knowledge graph extraction and modeling across multiple documents for health risk prediction . In Companion Proceedings of the ACM Web Conference 2024, pages 1182--1185

  3. [3]

    Ambedkar Dukkipati, Kawin Mayilvaghanan, Naveen Kumar Pallekonda, Sai Prakash Hadnoor, and Ranga Shaarad Ayyagari. 2025. https://arxiv.org/abs/2504.20058 Predictive ai with external knowledge infusion for stocks . arXiv preprint arXiv:2504.20058

  4. [4]

    Julia Gastinger, Timo Sztyler, Lokesh Sharma, Anett Schuelke, and Heiner Stuckenschmidt. 2023. https://dl.acm.org/doi/abs/10.1007/978-3-031-43418-1_32 Comparing apples and oranges? on the evaluation of methods for temporal knowledge graph forecasting . In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 533--549. Springer

  5. [5]

    Zhen Han, Peng Chen, Yunpu Ma, and Volker Tresp. 2020. https://openreview.net/forum?id=pGIHq1m7PU Explainable subgraph reasoning for forecasting on temporal knowledge graphs . In International Conference on Learning Representations

  6. [6]

    Zhen Han, Zifeng Ding, Yunpu Ma, Yujia Gu, and Volker Tresp. 2021. https://aclanthology.org/2021.emnlp-main.658/ Learning neural ordinary equations for forecasting future links on temporal knowledge graphs . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 8352--8364

  7. [7]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeGNn4cx5c Lora: Low-rank adaptation of large language models . In International Conference on Learning Representations (ICLR)

  8. [8]

    Yixin Ji, Juntao Li, Hai Ye, Kaixin Wu, Kai Yao, Jia Xu, Linjian Mo, and Min Zhang. 2025. https://doi.org/10.48550/arXiv.2501.02497 Test-time compute: from system-1 thinking to system-2 thinking . Preprint, arXiv:2501.02497

Show all 30 references
  1. [9]

    Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. 2020. https://aclanthology.org/2020.emnlp-main.541/ Recurrent event network: Autoregressive structure inference over temporal knowledge graphs . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Proc...

  2. [10]

    Dong-Ho Lee, Kian Ahrabian, Woojeong Jin, Fred Morstatter, and Jay Pujara. 2023. https://arxiv.org/abs/2305.10613 Temporal knowledge graph forecasting without knowledge using in-context learning . Preprint, arXiv:2305.10613

  3. [11]

    Kalev Leetaru and Philip A Schrodt. 2013. Gdelt: Global data on events, location, and tone, 1979--2012. In ISA annual convention, volume 2, pages 1--49. Citeseer

  4. [12]

    Aitor Lewkowycz, Anders Andreassen, David Dohan, Ethan Dyer, Henryk Michalewski, Vinay Ramasesh, Ambrose Slone, Cem Anil, Imanol Schlag, Theo Gutman-Solo, Yuhuai Wu, Behnam Neyshabur, Guy Gur-Ari, and Vedant Misra. 2022. https://arxiv.org/abs/2206.14858 Solving quantitative re...

  5. [13]

    Zixuan Li, Xiaolong Jin, Wei Li, Saiping Guan, Jiafeng Guo, Huawei Shen, Yuanzhuo Wang, and Xueqi Cheng. 2021. https://dl.acm.org/doi/10.1145/3404835.3462963 Temporal knowledge graph reasoning based on evolutional representation learning . In Proceedings of the 44th Internatio...

  6. [14]

    Ruotong Liao, Xu Jia, Yangzhe Li, Yunpu Ma, and Volker Tresp. 2024. https://arxiv.org/abs/2310.07793 Gentkg: Generative forecasting on temporal knowledge graph with large language models . Preprint, arXiv:2310.07793

  7. [15]

    Yushan Liu, Yunpu Ma, Marcel Hildebrandt, Mitchell Joblin, and Volker Tresp. 2022. https://cdn.aaai.org/ojs/20330/20330-13-24343-1-2-20220628.pdf Tlogic: Temporal logical rules for explainable link forecasting on temporal knowledge graphs . In Proceedings of the Thirty-Sixth A...

  8. [16]

    Ruilin Luo, Tianle Gu, Haoling Li, Junzhe Li, Zicheng Lin, Jiayi Li, and Yujiu Yang. 2024. https://arxiv.org/abs/2401.06072 Chain of history: Learning and forecasting with llms for temporal knowledge graph completion . Preprint, arXiv:2401.06072

  9. [17]

    Farzaneh Mahdisoltani, Joanna Biega, and Fabian M Suchanek. 2013. Yago3: A knowledge base from multilingual wikipedias. In CIDR

  10. [18]

    Shreyas Mangrulkar and 1 others. 2022. Peft: Parameter-efficient fine-tuning. https://github.com/huggingface/peft. GitHub repository, accessed May 2025

  11. [19]

    Meta AI . 2024. Meta llama 3: Open foundation and fine-tuned chat models. https://ai.meta.com/blog/meta-llama-3/. Accessed: 2025-05-16

  12. [20]

    all-mpnet-base-v2

    Sentence-Transformers. all-mpnet-base-v2. https://huggingface.co/sentence-transformers/all-mpnet-base-v2. Accessed: 2025-05-19

  13. [21]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. https://arxiv.org/abs/2408.03314 Scaling llm test-time compute optimally can be more effective than scaling model parameters . Preprint, arXiv:2408.03314

  14. [22]

    Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. https://proceedings.neurips.cc/paper/2020/file/c3a690be93aa602ee2dc0ccab5b7b67e-Paper.pdf Mpnet: Masked and permuted pre-training for language understanding . In Advances in Neural Information Processing Systems...

  15. [23]

    Haohai Sun, Jialun Zhong, Yunpu Ma, Zhen Han, and Kun He. 2021. https://arxiv.org/abs/2109.04101 Timetraveler: Reinforcement learning for temporal knowledge graph forecasting . arXiv preprint arXiv:2109.04101

  16. [24]

    Hugo Touvron, Louis Martin, Kevin Stone, Abdullah Al-Dujaili, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, and 1 others. 2023. https://arxiv.org/abs/2307.09288 Llama 2: Open foundation and fine-tuned chat models . arXiv preprint arXiv:2307.09288

  17. [25]

    Volker Tresp, Crist \'o bal Esteban, Yinchong Yang, Stephan Baier, and Denis Krompa . 2015. https://arxiv.org/abs/1511.07972 Learning with memory embeddings . arXiv preprint arXiv:1511.07972

  18. [26]

    Rakshit Trivedi, Hanjun Dai, Yichen Wang, and Le Song. 2017. https://proceedings.mlr.press/v70/trivedi17a.html Know-evolve: Deep temporal reasoning for dynamic knowledge graphs . In Proceedings of the 34th International Conference on Machine Learning, pages 3462--3471. PMLR

  19. [27]

    O mer Faruk Akg \

    Shangshang Wang, Julian Asilis, \"O mer Faruk Akg \"u l, Enes Burak Bilgin, Ollie Liu, and Willie Neiswanger. 2025. https://arxiv.org/abs/2504.15777 Tina: Tiny reasoning models via lora . arXiv preprint arXiv:2504.15777

  20. [28]

    Yuwei Xia, Ding Wang, Qiang Liu, Liang Wang, Shu Wu, and Xiao-Yu Zhang. 2024. https://doi.org/10.18653/v1/2024.findings-acl.955 Chain-of-history reasoning for temporal knowledge graph forecasting . In Findings of the Association for Computational Linguistics: ACL 2024, pages 1...

  21. [29]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  22. [30]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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