Pith. sign in

REVIEW 3 major objections 5 minor 44 references

A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning

T0 review · 3 major / 5 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 DGAR claims that generating and adaptively replaying an entity's historical distribution, guided by the current reasoning model, reduces catastrophic forgetting in temporal knowledge graph reasoning better than replaying past facts alone.

desk verdict New combination of context-prompt replay and diffusion-guided generation for continual TKGR, with real experimental breadth, but the 'significantly outperforms' claim lacks statistical support. read the letter →

arxiv 2506.04083 v1 pith:47K4PASX submitted 2025-06-04 cs.IR

Zhiyu Zhang , Wei Chen , Youfang Lin , Huaiyu Wan This is my paper
classification cs.IR
keywords continuallearningtemporalknowledgegraphreasoningcatastrophicforgettinggenerativereplaydiffusionmodelsembeddingsadaptivedistributionshift
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 tries to solve catastrophic forgetting in temporal knowledge graph reasoning when new facts arrive over time. Instead of replaying individual old facts, it proposes DGAR, which generates a representation of each entity's historical distribution from the full historical context, guided by the current reasoning model to keep the generated past compatible with the present. The generated history is injected layer by layer into the current entity representations with an adaptive weight. The authors report that this beats fine-tuning, experience replay, and prior continual-learning baselines, with the largest stated gains on historical tasks, and claim the method transfers to different GNN-based reasoners.

What carries the argument

The load-bearing pieces are Historical Context Prompts (HCPs), Diffusion-Enhanced Historical Distribution Generation (Diff-HDG), and Deep Adaptive Replay (DAR). HCPs are sampling units: for a queried entity, the set of triples involving that entity at a sampled historical timestamp, preserving context rather than isolated facts. Diff-HDG uses a pre-trained diffusion model, conditioned on the historical neighbor and relation embeddings, to denoise an entity noise vector into a historical distribution representation; during denoising it takes gradient steps from the current TKGR model's softmax score on historical facts, which is where the common-feature enhancement happens. DAR injects the generated historical representation into the current representation at every layer of the base GNN reasoner with a learned adaptive weight alpha, which is what balances old and new knowledge.

What would settle it

Build two synthetic TKG streams with identical facts but different ordering, one where neighbors change smoothly between adjacent timestamps and one where each timestamp completely rewrites the entities' neighbors; if DGAR's historical-task MRR does not drop in the abrupt stream relative to the smooth stream, the minor-distribution-difference assumption is not doing the claimed work.

Watch

Extended reading notes

Core claim

DGAR's central claim is that catastrophic forgetting in temporal knowledge graphs can be reduced by replaying a generated distribution rather than replaying facts. For each entity touched by a new query, the model builds historical context prompts from a sample of past timestamps, feeds them into a pre-trained denoising diffusion model to generate the entity's historical distribution representation, and uses the current task's model to amplify features shared between historical and current distributions while suppressing conflicting ones. A layer-by-layer deep adaptive replay then mixes this generated historical representation into the current representation at each evolution unit using a scalar alpha. The authors report average MRR improvements of 4.01% on the current task and 8.23% on historical tasks over the strongest baselines, with larger Hits@10 gains on historical tasks.

Load-bearing premise

The method assumes that adjacent timestamps in a temporal knowledge graph differ only slightly, so the previous moment's trained model can stand in for the current moment's model when steering the generated historical replay.

Editorial extensions

If this is right

  • Continual learners for TKGs can keep historical knowledge without storing all raw history, since only sampled context prompts and a diffusion model are needed for replay.
  • The method generalizes beyond RE-GCN: the paper reports it also improves TiRGN and LogCL over fine-tuning under the continual-learning setting.
  • The adaptive replay weight provides a tunable trade-off between preserving old distributions and fitting new ones; ablations show removing it lowers both current and historical MRR.
  • Because random selection of historical context prompts beats selecting the nearest k timestamps, replay benefits from diversity across history, not just recency.

Reading between the lines

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

  • An implication the authors leave implicit: the diffusion-guider step is essentially a compatibility filter, similar to gradient-based regularization, so a cheaper approximation using one gradient step on a contrastive loss might reproduce part of the gain.
  • Because the method only models entity distributions and leaves relations mostly static, its advantage should shrink on TKGs where relation semantics drift as much as entity semantics; this is testable on datasets with strong relation turnover.
  • The gap between current-task and historical-task gains suggests DGAR may help mainly by regularizing rather than by improving the current task; holding current-task accuracy fixed would separate these two effects.
  • In a streaming deployment the pre-trained diffusion model itself is continually fine-tuned, so its own forgetting could compound the method's errors; the paper does not measure the diffusion model's memory separately.
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 / 5 minor

Summary. The paper proposes DGAR, a continual learning method for temporal knowledge graph reasoning. It introduces Historical Context Prompts (HCPs) as sampling units, a diffusion-based generator (Diff-HDG) that produces historical entity distribution representations under guidance from the current TKGR model, and a Deep Adaptive Replay (DAR) mechanism that injects these representations into the base model at multiple layers. Experiments on ICE14, ICE18, ICE05-15, and GDELT report MRR and Hits@k improvements over fine-tuning and several continual learning baselines, with additional ablations, efficiency analysis, and an extension to different base models.

Significance. If validated, the paper addresses a genuine and current problem: preserving historical context and mitigating distribution conflicts in continual learning for TKGR. The idea of generating replay representations with a pretrained diffusion model rather than replaying raw facts is a novel direction, and the paper includes useful ablations, an efficiency comparison, and code availability. The main contribution is plausible, but the central claim of significant improvement currently rests on point estimates without variance or significance testing, and the guidance mechanism's self-referential nature needs explicit scrutiny.

major comments (3)
  1. [Section 5.2, Tables 1 and 2] The paper reports that each dataset is tested five times and average results are reported, but no standard deviations, confidence intervals, or significance tests are given for any of the main tables. The abstract's claim that DGAR 'significantly outperforms baselines' is a statistical assertion that is not supported by statistical evidence. Given the heterogeneous gains (e.g., GDELT MRR 23.25 vs. 15.56 for TIE, roughly a 49% relative improvement, versus ICE18 MRR 36.53 vs. 34.45, roughly 6%) and per-dataset tuning of the hyperparameter k, seed sensitivity and tuning choices could account for part of the reported differences. Please report variance or confidence intervals for the five runs and perform pairwise significance tests for the key comparisons against the strongest baseline.
  2. [Section 4.2, Eqs. (9)-(10)] The guidance step applies gradient ascent on the current TKGR model's softmax score of historical facts to modify the generated historical representation. Subsequently, Section 4.4's replay loss L_{t,r} trains the current model on the same historical facts. This creates a self-referential loop: the generated historical distribution is shaped by the current model's inductive biases and then used to train that same model, which could reinforce existing biases rather than recovering true historical knowledge. The paper's qualitative description that common features are 'enhanced' and conflicting features 'weakened' is not a formal justification. Please provide a theoretical or empirical demonstration that this guidance does not merely fit the current model to itself, for example by comparing against unguided diffusion or against generation guided by a frozen earlier model.
  3. [Section 4.2] The approximation of the current parameters θ_t by the previous step's θ_{t-1} is justified by the assertion that 'adjacent timestamps in TKGs show only minor distribution differences.' This assumption is not validated empirically, and it is particularly questionable for GDELT, which has 15-minute granularity. If the distribution shift between adjacent timestamps is large, the guidance could distort the generated historical representations, and the subsequent replay could reinforce outdated model behavior. Please add a quantitative analysis of distribution shift between consecutive timestamps (e.g., using the distributional distance of entity representations) or a sensitivity analysis in which guidance uses models from different time lags, to assess the robustness of this key approximation.
minor comments (5)
  1. [Section 4.2, Eq. (6)] Equation (6) writes the generator as p_φ(X_n, f_{θ_t}, Prompt_replay), but the text explains that θ_t is approximated by θ_{t-1}; the notation should be made consistent, for example by writing f_{θ_{t-1}} explicitly.
  2. [Section 5.1] The metric 'Average' is described as the 'average performance across all previous test sets,' but the text should also clarify whether the current test set is included or excluded, and how this is computed when there are many previous tasks.
  3. [Appendix B.5, Table 5] The time comparison between Retrain and DGAR is not apples-to-apples: retraining processes the entire dataset at each task, while DGAR processes only the new task plus replay. This difference should be stated directly in the table caption or surrounding text to avoid misleading readers.
  4. [Appendix A.2] The pretraining of the diffusion model via continual learning and the role of Eq. (16) in preserving the DM's historical knowledge are only briefly described. A short pseudocode or a more detailed explanation of how the DM is updated without forgetting would improve reproducibility.
  5. [Figures 1 and 5] The UMAP visualization is qualitative and depends on the chosen random stage. I recommend reporting a quantitative measure of distribution overlap or separation (e.g., maximum mean discrepancy or Bhattacharyya distance) to support the claim that DGAR resolves distribution conflicts.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: DGAR's generation-replay loop is a regularization mechanism, and the reported gains are empirical comparisons against external baselines, not derivations equivalent to the method's inputs.

full rationale

Walked the derivation chain: HCPs are sampled from historical facts; Diff-HDG generates entity representations by denoising conditioned on historical neighbors and relations and applies gradient guidance from the current TKGR model; DAR injects these representations; training minimizes a current-task loss plus a replay loss on historical facts. The guidance step uses the model's own scores to shape generated representations, but this is not a prediction claim: no result is derived from the guidance by construction, and final MRR/Hits@k are measured on held-out current and historical test sets against external baselines (FT, ER, TIE, LKGE, IncDE). The replay loss L_{t,r} additionally trains on the original historical facts, so the loop has external grounding. Self-citations (Chen et al. 2024b/2024c, Wu et al. 2023, Zhao et al. 2025) are contextual or serve as an additional base model in Appendix B.8; none carries the central argument. The stated approximation theta_t approximated by theta_{t-1} is an explicit modeling assumption, not a circular reduction. The lack of variance or significance reporting is an evidentiary weakness of the 'significantly outperforms' claim, but it is not circularity under the review criteria.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method introduces new data structures (historical context prompts, generated historical distributions, adaptive fusion) but no new physical or conceptual entities. The central claim rests on several domain assumptions about temporal distribution smoothness and the faithfulness of diffusion-generated representations, plus a set of tuned hyperparameters (k, γ, α, µ, L).

free parameters (5)
  • k (number of sampled historical time slices for HCP) = 35 (ICE14), 25 (ICE18), 40 (ICE05-15), 32 (GDELT)
    Tuned per dataset (Appendix B.3); sensitivity analysis (B.4) shows it affects historical task MRR.
  • γ (guidance strength in Diff-HDG) = 1
    Set as optimal coefficient (Appendix B.3).
  • α (adaptive fusion weight in DAR) = learned, not explicitly reported
    A scalar that balances historical and current representations (Eq. 12); unclear if shared or per-layer, and the value is not reported.
  • µ (replay loss coefficient) = 1
    Set to 1 (Section 4.4 and B.3).
  • L (number of DAR layers) = 3
    Set as optimal (Appendix B.3).
assumptions (5)
  • domain assumption Adjacent timestamps in TKGs have minor distribution differences, allowing θ_{t-1} to approximate θ_t for guidance.
    Invoked in Section 4.2 to justify using the previous model for guidance before the current model is trained.
  • domain assumption Historical semantics of an entity are captured by the set of triples involving it at past timesteps.
    Basis for HCP construction (Eq. 4).
  • domain assumption A pre-trained diffusion model can generate meaningful historical entity distributions from conditioned prompts.
    Core to Diff-HDG; empirically evaluated but not formally justified.
  • ad hoc to paper Gradient ascent on the current model's softmax score of historical facts enhances common features and weakens conflicting features.
    Proposed in Eqs. 9-10 without a formal proof or measurement of the claimed feature effects.
  • standard math Standard DDPM training and sampling assumptions hold for entity embedding space.
    Uses standard diffusion equations (Eqs. 1-3).

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning." pith.science (2026). https://pith.science/paper/47K4PASX

@misc{pith2026250604083,
  author       = {Pith},
  title        = {Pith review of: A Generative Adaptive Replay Continual Learning Model for Temporal Knowledge Graph Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/47K4PASX}},
  note         = {Machine review of arXiv:2506.04083}
}
read the original abstract

Recent Continual Learning (CL)-based Temporal Knowledge Graph Reasoning (TKGR) methods focus on significantly reducing computational cost and mitigating catastrophic forgetting caused by fine-tuning models with new data. However, existing CL-based TKGR methods still face two key limitations: (1) They usually one-sidedly reorganize individual historical facts, while overlooking the historical context essential for accurately understanding the historical semantics of these facts; (2) They preserve historical knowledge by simply replaying historical facts, while ignoring the potential conflicts between historical and emerging facts. In this paper, we propose a Deep Generative Adaptive Replay (DGAR) method, which can generate and adaptively replay historical entity distribution representations from the whole historical context. To address the first challenge, historical context prompts as sampling units are built to preserve the whole historical context information. To overcome the second challenge, a pre-trained diffusion model is adopted to generate the historical distribution. During the generation process, the common features between the historical and current distributions are enhanced under the guidance of the TKGR model. In addition, a layer-by-layer adaptive replay mechanism is designed to effectively integrate historical and current distributions. Experimental results demonstrate that DGAR significantly outperforms baselines in reasoning and mitigating forgetting.

Figures

Figures reproduced from arXiv: 2506.04083 by the authors.

Figure 1
Figure 1. The distributions of the same set of entity [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture diagram of DGAR. Following the CL paradigm, each snapshot of the TKGs is [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Performance of different base TKGR models. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Effect of memorizing old knowledge in CL. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization case study of entity distribution. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]
Figure 6
Figure 6. Figure 6: Sensitivity Analysis. conducted on NVIDIA A40. B.4 Sensitivity Analysis In this section, we conduct experiments on ICE14 and ICE18 to further analyze the impact of hy￾perparameter k in DGAR. The hyperparameter k means the replay data consists of HCPs at k differ￾ent ti…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 30 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. 2021. Structured denoising diffusion models in discrete state-spaces. Advances in Neural Information Processing Systems, 34:17981--17993

    work page 2021

  4. [4]

    Yuxiang Cai, Qiao Liu, Yanglei Gan, Changlin Li, Xueyi Liu, Run Lin, Da Luo, and JiayeYang JiayeYang. 2024. Predicting the unpredictable: Uncertainty-aware reasoning over temporal knowledge graphs via diffusion process. In Findings of the Association for Computational Linguistics ACL 2024, pages 5766--5778

    work page 2024

  5. [5]

    Kai Chen, Ye Wang, Yitong Li, Aiping Li, Han Yu, and Xin Song. 2024 a . https://doi.org/10.18653/v1/2024.acl-long.8 A unified temporal knowledge graph reasoning model towards interpolation and extrapolation . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 117--132, Bangkok, Thailan...

    work page doi:10.18653/v1/2024.acl-long.8 2024

  6. [6]

    Wei Chen, Huaiyu Wan, Yuting Wu, Shuyuan Zhao, Jiayaqi Cheng, Yuxin Li, and Youfang Lin. 2024 b . https://doi.org/10.1109/ICDE60146.2024.00062 Local-global history-aware contrastive learning for temporal knowledge graph reasoning . In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pages 733--746

    arXiv 2024

  7. [7]

    Wei Chen, Yuting Wu, Shuhan Wu, Zhiyu Zhang, Mengqi Liao, Youfang Lin, and Huaiyu Wan. 2024 c . https://api.semanticscholar.org/CorpusID:274982454 Cogntke: A cognitive temporal knowledge extrapolation framework . ArXiv, abs/2412.16557

    work page Pith review arXiv 2024

  8. [8]

    Yubo Chen, Shaoru Guo, Kang Liu, and Jun Zhao. 2023. Large language models and knowledge graphs. In Proceedings of the 22nd Chinese National Conference on Computational Linguistics (Volume 2: Frontier Forum), pages 67--76

    work page 2023

Show all 44 references
  1. [9]

    Yuanning Cui, Yuxin Wang, Zequn Sun, Wenqiang Liu, Yiqiao Jiang, Kexin Han, and Wei Hu. 2023. Lifelong embedding learning and transfer for growing knowledge graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 4217--4224

    2023

  2. [10]

    Shib Sankar Dasgupta, Swayambhu Nath Ray, and Partha Talukdar. 2018. Hyte: Hyperplane-based temporally aware knowledge graph embedding. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2001--2011

    2018

  3. [11]

    Rishab Goel, Seyed Mehran Kazemi, Marcus Brubaker, and Pascal Poupart. 2020. Diachronic embedding for temporal knowledge graph completion. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 3988--3995

    2020

  4. [12]

    Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. 2022. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933

    2022 arXiv

  5. [13]

    Saiping Guan, Xueqi Cheng, Long Bai, Fujun Zhang, Zixuan Li, Yutao Zeng, Xiaolong Jin, and Jiafeng Guo. 2022. What is event knowledge graph: A survey. IEEE Transactions on Knowledge and Data Engineering, 35(7):7569--7589

    2022

  6. [14]

    Bernal Jim \'e nez Guti \'e rrez, Yiheng Shu, Yu Gu, Michihiro Yasunaga, and Yu Su. 2024. Hipporag: Neurobiologically inspired long-term memory for large language models. arXiv preprint arXiv:2405.14831

    2024 arXiv

  7. [15]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840--6851

    2020

  8. [16]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. 2020. Heterogeneous graph transformer. In Proceedings of the web conference 2020, pages 2704--2710

    2020

  9. [17]

    Rikui Huang, Wei Wei, Xiaoye Qu, Shengzhe Zhang, Dangyang Chen, and Yu Cheng. 2024. https://doi.org/10.18653/v1/2024.acl-long.580 Confidence is not timeless: Modeling temporal validity for rule-based temporal knowledge graph forecasting . In Proceedings of the 62nd Annual Meet...

    2024 doi

  10. [18]

    Woojeong Jin, Meng Qu, Xisen Jin, and Xiang Ren. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.541 Recurrent event network: Autoregressive structure inferenceover temporal knowledge graphs . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Pro...

    2020 doi

  11. [19]

    Timoth \' e e Lacroix, Guillaume Obozinski, and Nicolas Usunier. 2020. Tensor decompositions for temporal knowledge base completion. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020 . OpenReview.net

    2020

  12. [20]

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

    2018

  13. [21]

    Eunhae Lee. 2024. https://arxiv.org/abs/2407.00176 The impact of model size on catastrophic forgetting in online continual learning . Preprint, arXiv:2407.00176

    2024 arXiv

  14. [22]

    Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. 2022 a . Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328--4343

    2022

  15. [23]

    Yujia Li, Shiliang Sun, and Jing Zhao. 2022 b . Tirgn: Time-guided recurrent graph network with local-global historical patterns for temporal knowledge graph reasoning. In IJCAI, pages 2152--2158

    2022

  16. [24]

    Yujia Li, Shiliang Sun, and Jing Zhao. 2022 c . Tirgn: Time-guided recurrent graph network with local-global historical patterns for temporal knowledge graph reasoning. In Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence, IJCAI 2022, Vi...

    2022

  17. [25]

    Zixuan Li, Xiaolong Jin, Wei Li, Saiping Guan, Jiafeng Guo, Huawei Shen, Yuanzhuo Wang, and Xueqi Cheng. 2021. Temporal knowledge graph reasoning based on evolutional representation learning

    2021

  18. [26]

    Sen Lin, Li Yang, Deliang Fan, and Junshan Zhang. 2022. Beyond not-forgetting: Continual learning with backward knowledge transfer. Advances in Neural Information Processing Systems, 35:16165--16177

    2022

  19. [27]

    Jiajun Liu, Wenjun Ke, Peng Wang, Ziyu Shang, Jinhua Gao, Guozheng Li, Ke Ji, and Yanhe Liu. 2024 a . Towards continual knowledge graph embedding via incremental distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 8759--8768

    2024

  20. [28]

    Jiajun Liu, Wenjun Ke, Peng Wang, Jiahao Wang, Jinhua Gao, Ziyu Shang, Guozheng Li, Zijie Xu, Ke Ji, and Yining Li. 2024 b . Fast and continual knowledge graph embedding via incremental lora. arXiv preprint arXiv:2407.05705

    2024 arXiv

  21. [29]

    Xiao Long, Liansheng Zhuang, Aodi Li, Houqiang Li, and Shafei Wang. 2024. Fact embedding through diffusion model for knowledge graph completion. In Proceedings of the ACM on Web Conference 2024, pages 2020--2029

    2024

  22. [30]

    Mehrnoosh Mirtaheri, Mohammad Rostami, and Aram Galstyan. 2023. History repeats: Overcoming catastrophic forgetting for event-centric temporal knowledge graph completion. arXiv preprint arXiv:2305.18675

    2023 arXiv

  23. [31]

    David Rolnick, Arun Ahuja, Jonathan Schwarz, Timothy Lillicrap, and Gregory Wayne. 2019. Experience replay for continual learning. Advances in neural information processing systems, 32

    2019

  24. [32]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684--10695

    2022

  25. [33]

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. 2015. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256--2265. PMLR

    2015

  26. [34]

    Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. 2023. Sketch-guided text-to-image diffusion models. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1--11

    2023

  27. [35]

    Jiapu Wang, Kai Sun, Linhao Luo, Wei Wei, Yongli Hu, Alan Wee-Chung Liew, Shirui Pan, and Baocai Yin. 2024. Large language models-guided dynamic adaptation for temporal knowledge graph reasoning. arXiv preprint arXiv:2405.14170

    2024 arXiv

  28. [36]

    Jiapeng Wu, Yishi Xu, Yingxue Zhang, Chen Ma, Mark Coates, and Jackie Chi Kit Cheung. 2021. Tie: A framework for embedding-based incremental temporal knowledge graph completion. In Proceedings of the 44th international ACM SIGIR conference on research and development in inform...

    2021

  29. [37]

    Shuhan Wu, Huaiyu Wan, Wei Chen, Yuting Wu, Junfeng Shen, and Youfang Lin. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.676 Towards enhancing relational rules for knowledge graph link prediction . In Findings of the Association for Computational Linguistics: EMNLP 202...

    2023 doi

  30. [38]

    Yujie Xing, Xiao Wang, Yibo Li, Hai Huang, and Chuan Shi. 2024. Less is more: on the over-globalizing problem in graph transformers. arXiv preprint arXiv:2405.01102

    2024 arXiv

  31. [39]

    Yi Xu, Junjie Ou, Hui Xu, and Luoyi Fu. 2023 a . Temporal knowledge graph reasoning with historical contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 4765--4773

    2023

  32. [40]

    Yi Xu, Junjie Ou, Hui Xu, and Luoyi Fu. 2023 b . Temporal knowledge graph reasoning with historical contrastive learning. In AAAI

    2023

  33. [41]

    Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. 2023. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 142...

    2023

  34. [42]

    Jiasheng Zhang, Jie Shao, and Bin Cui. 2023 a . Streame: Learning to update representations for temporal knowledge graphs in streaming scenarios. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 622--631

    2023

  35. [43]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023 b . Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836--3847

    2023

  36. [44]

    Shuyuan Zhao, Wei Chen, Boyan Shi, Liyong Zhou, Shuohao Lin, and Huaiyu Wan. 2025. https://doi.org/10.1609/AAAI.V39I12.33459 Spatial-temporal knowledge distillation for takeaway recommendation . In AAAI-25, Sponsored by the Association for the Advancement of Artificial Intelli...

    2025 doi

Pith tools

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