Pith. sign in

REVIEW 3 major objections 5 minor 56 references

This paper claims that LLM-guided cluster-consensus OOD exposure, injected as a score-level regularizer, consistently improves topology-driven out-of-distribution detection on text-attributed graphs, cutting FPR95 by at least 7% and outperf

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

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

T0 review · deepseek-v4-flash

2026-08-03 00:01 UTC pith:7OF2JKBI

load-bearing objection Plausible, useful plug-in for TAG OOD detection; the headline FPR95 gains need error bars and an exposure-purity measurement before I'd trust them. the 3 major comments →

arxiv 2602.11641 v2 pith:7OF2JKBI submitted 2026-02-12 cs.LG

Both Topology and Text Matter: Revisiting LLM-guided Out-of-Distribution Detection on Text-attributed Graphs

Yinlin Zhu , Di Wu , Xu Wang , Guocong Quan , Miao Hu This is my paper
classification cs.LG
keywords out-of-distribution detectiontext-attributed graphsgraph neural networkslarge language modelsOOD exposureconsensus filteringscore regularizationplug-and-play
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that on text-attributed graphs — networks where each node carries a piece of text — out-of-distribution detection is best done by combining two sources of evidence: the graph structure, which existing topology-driven detectors already exploit, and the semantics of node texts, which large language models can judge. It proposes LG-Plug, a plug-and-play module that first aligns graph and text embeddings, then asks an LLM to label small batches of near-centroid unlabeled nodes cluster by cluster, keeping only clusters where the LLM's top categories are all outside the in-distribution label set. Those consensus-validated OOD nodes become a regularization signal that pushes a detector's OOD scores apart from in-distribution scores, without changing the detector's architecture. The paper reports that this consistently improves topology-driven detectors (≥7% FPR95 reduction) and outperforms LLM-based baselines (≥5% FPR95 reduction) across six TAG benchmarks, resolving the reliability-informativeness trade-off that plagues earlier LLM-based OOD prior generation.

Core claim

The central claim is that LLM-derived OOD exposure is most effective when it is (a) harvested at the cluster level from unlabeled nodes in a jointly learned topology-text embedding space, so that consensus among multiple nodes filters out ambiguous annotations, and (b) injected as a margin-based score regularizer (Eq. 15) that is agnostic to the detector's internals, rather than as a bespoke training signal. The paper demonstrates on six TAG datasets that integrating LG-Plug with representative topology-driven detectors yields at least a 7% reduction in FPR95, and that it beats three LLM-based baselines by at least 5% FPR95, while cutting LLM query cost to O(M·b·T), independent of graph size

What carries the argument

The object that carries the argument is the OOD exposure set V_exp, constructed in three steps: (1) a topology-text representation alignment that trains a GCN graph encoder and a Transformer text encoder with node-level and edge-level contrastive losses, producing embeddings in which ID and OOD nodes are roughly separated; (2) K-means clustering over unlabeled node embeddings, selection of near-centroid nodes per cluster, and iterative LLM prompting with a dynamic category codebook until a batch reaches consensus — the top-K consensus filter (Eqs. 12–13) admits a cluster's nodes into V_exp only if the top categories contain no ID class; and (3) a margin-based ReLU regularization term (Eq. 15

Load-bearing premise

The LLM's category assignments, after the top-K consensus filter, reliably separate true OOD nodes from unlabeled in-distribution nodes — if the LLM systematically assigns novel categories to a meaningful share of ID texts, the OOD exposure set is polluted and the regularizer would push in-distribution scores upward, degrading detection.

What would settle it

A direct test: on a dataset with ground-truth OOD labels, measure the fraction of ID nodes in the exposure set V_exp. If that ID contamination rate is high (e.g., above the LLM's baseline misassignment rate) yet FPR95 still improves by 7%, the mechanism is not doing what the paper says — or the gains come from something other than reliable exposure.

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

If this is right

  • Existing topology-driven graph OOD detectors can be upgraded with LLM semantics by adding one regularization term, with no architectural change — the paper shows this for representative energy- and score-propagation detectors.
  • LLM query cost is decoupled from graph size (O(M·b·T) instead of O(N_s·T)), so the approach remains practical for large graphs such as ogbn-arxiv.
  • The reliability-informativeness trade-off in LLM-generated OOD priors is mitigated by cluster-level consensus rather than by more elaborate prompting alone.
  • Jointly modeling text and topology gives larger detection gains than either modality alone, suggesting that both signals are complementary for TAG OOD detection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The hard exclusion rule (top-K categories disjoint from ID labels) may be too brittle when the ID label space is large or categories overlap; a calibrated confidence threshold on the LLM's assignments could make the exposure set more robust without sacrificing precision.
  • The approach implicitly assumes the unlabeled pool contains a non-negligible fraction of OOD nodes; on graphs where OOD is very rare, cluster-level consensus may yield few exposures, and the regularizer's benefit could shrink.
  • Because the text encoder is trained from scratch (63M parameters) without pretrained weights, using a pretrained LLM-based encoder could plausibly further improve alignment quality, at the cost of the reported efficiency.
  • The same cluster-consensus mechanism could be adapted to open-world or dynamic graph settings, where new semantic clusters emerge over time — the paper lists this as future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes LG-Plug, a plug-and-play method for OOD detection on text-attributed graphs. It first aligns a GNN encoder and a text encoder via node- and edge-level contrastive losses, then clusters unlabeled nodes and uses iterative LLM prompting with a category codebook to identify a consensus-driven OOD exposure set. This exposure set is used to add a margin-based score regularization to existing topology-driven detectors such as GNNSafe and GRASP. Experiments on six TAG benchmarks report consistent FPR95 reductions of at least 7% over topology baselines and at least 5% over LLM-based baselines, along with ablations, sensitivity analyses, efficiency comparisons, and word-cloud visualizations of OOD exposure quality.

Significance. If the reported gains hold under rigorous statistical and leakage-controlled evaluation, LG-Plug would be a useful contribution: it is detector-agnostic, combines topology and LLM semantics, and is relatively lightweight in LLM queries. The framework is clearly motivated and the implementation details are concrete, with code and prompts provided. However, the central empirical claim is currently under-supported: Tables 1–2 show only point estimates despite Appendix E stating that mean and variance over runs were computed, and the method uses all unlabeled test nodes, including OOD nodes, to construct the exposure set that regularizes training. The purity of the exposure set is also validated only qualitatively. These issues make the magnitude and even the direction of the reported improvements uncertain.

major comments (3)
  1. [Appendix E; Tables 1–2] Appendix E states that all experiments were run without fixing random seeds and results are reported as mean and variance over multiple standardized runs, but Tables 1 and 2 contain only point estimates. No standard deviations, confidence intervals, or significance tests are provided. The paper's central claim of a consistent ≥7% FPR95 reduction cannot be assessed for statistical reliability. Please report the variance and run the appropriate significance tests, or state clearly if the tables are single-run results.
  2. [§4.2–4.3, Eqs. (13) and (15); Appendix A] The exposure set V_exp is built from unlabeled nodes V_u, and Appendix A states that all OOD nodes are placed exclusively in the test set. Thus, in this transductive setup, V_exp can contain the very OOD test nodes whose scores are later evaluated. Eq. (13) adds those nodes to V_exp, and Eq. (15) regularizes the detector by pushing their scores above Δ2. This leaks test-set OOD membership into training and can inflate the reported FPR95/AUROC gains. To support the claimed advantage, evaluate in an inductive setting or hold out a portion of OOD nodes that are never used in V_exp or the regularizer, and report performance on that held-out set.
  3. [§4.2, Eqs. (10)–(13); Fig. 5] The consensus filter checks only whether the top-K codebook categories are disjoint from Y_id (Eq. 12), not whether individual nodes assigned to those categories are genuinely OOD. Eq. (10) can assign an ID text to a newly created category when the text does not exactly match an ID class name, and Eq. (13) then places the node in V_exp. If such ID contamination occurs, Eq. (15) would push ID scores upward and could offset the reported FPR95 reduction. The only evidence of V_exp purity is the qualitative word cloud in Fig. 5. Please report quantitative precision/recall of V_exp with respect to ground-truth OOD membership per dataset, and show how performance varies with contamination rate.
minor comments (5)
  1. [Eq. (2)] The sentence after Eq. (2) says 'θ_GE signifies the parameter set of the text encoder'; this should be θ_TE. Please fix the typo.
  2. [Table 2] The rows after the lower bound are difficult to parse: all three intermediate rows show the same two checkmarks, making it unclear which components are ablated. Please label each row explicitly or use a dash for removed components.
  3. [Algorithm 1] In line 17, 'LLM trail' should be 'LLM trial'. Please correct the typo.
  4. [References] GLIP-OOD is cited twice as [40] and [41]; these are the same paper. Please deduplicate.
  5. [Abstract/Conclusion] The phrase 'LG-Plug avoids reduces time cost of LLM querying' is grammatically awkward. Please rephrase.

Circularity Check

0 steps flagged

No significant circularity; transductive exposure is an explicit protocol choice, not a derivation loop.

full rationale

The claimed derivation chain—topology-text alignment (Eqs. 1–6), clustering (Eqs. 7–8), LLM annotation (Eqs. 9–13), and score regularization (Eqs. 14–16)—does not contain any step in which an output is defined to be equal to an input or in which a fitted quantity is renamed as a prediction. The OOD exposure set V_exp is generated by LLM consensus and used only as a score-level regularizer; the FPR95/AUROC improvements in Table 1 are empirical comparisons against external topology-driven and LLM-based baselines. The semi-supervised transductive protocol is explicit: Sec. 3 states that unlabeled nodes V_u comprise both ID and OOD nodes, so using unlabeled (including test) node texts for exposure is part of the intended setting and is also consistent with the baselines' use of the graph. The strongest caveat is that V_exp purity is not quantitatively verified—Eqs. 12–13 only check cluster-level label disjointness, and Fig. 5's word clouds do not provide precision/recall. This is a correctness and robustness risk: if ID nodes are assigned novel LLM categories, the reported gains would be weakened or reversed, not forced by construction. No load-bearing self-citations were found. Therefore the derivation is not circular.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The method rests on several untested assumptions about LLM reliability, cluster semantics, and transfer of margin regularization. The central hyperparameters are chosen by search and not all values are disclosed. No invented physical or scientific entities are introduced.

free parameters (9)
  • λ (alignment trade-off) = searched over {0.01,0.1,1,10}
    Balances node-level and edge-level alignment losses in Eq. (6); chosen per dataset.
  • τ (similarity temperature) = not reported
    Scales the node similarity matrix Λ1 in Eq. (3); introduced ad hoc.
  • M (number of clusters) = searched over {5,10,15,20,40,50,100}
    K-means granularity in Eq. (7); primary sensitivity axis.
  • ρ (intra-cluster filtering ratio) = searched in [0,1]
    Selects near-centroid representative subset in Eq. (8).
  • b (LLM query batch size) = searched over {1,2,3,4,5}
    Number of texts queried per consensus batch.
  • T (LLM query trials) = fixed to 3
    Termination cap for sequential codebook updates.
  • K (codebook top categories) = fixed to 2
    Filters codebook to most frequent categories in Eq. (12).
  • Δ1/Δ2 (score margins) = not specified
    ID and OOD margins in the regularization loss Eq. (15).
  • β (regularization weight) = not reported
    Trade-off between supervised loss and OOD-exposure regularization in Eq. (16).
axioms (5)
  • domain assumption Unlabeled nodes partition into ID nodes and OOD nodes with disjoint label spaces.
    Sec. 3: V_u = V_u_id ∪ V_u_ood and Y_ood is unknown.
  • domain assumption LLM category assignments on node text are reliable enough that cluster-level consensus identifies true OOD nodes.
    Eqs. (10)-(13): if the LLM systematically assigns ID texts to novel categories, V_exp is polluted.
  • domain assumption K-means clusters in the aligned embedding space group semantically coherent nodes.
    Eqs. (7)-(8): near-centroid representatives are assumed to share category semantics.
  • domain assumption Score regularization on exposed nodes transfers to unexposed OOD nodes.
    Eqs. (15)-(16): margins imposed on V_exp are assumed to separate the whole ID/OOD score distribution.
  • domain assumption The contrastive alignment losses can be jointly optimized to produce fine-grained embeddings without labeled OOD data.
    Eqs. (4)-(6): L_node and L_edge are trained from scratch; no guarantee of cluster separability.

pith-pipeline@v1.3.0-alltime-deepseek · 19212 in / 14217 out tokens · 111797 ms · 2026-08-03T00:01:35.710054+00:00 · methodology

0 comments
read the original abstract

Text-attributed graphs (TAGs) associate nodes with textual attributes and graph structure, enabling GNNs to jointly model semantic and structural information. Although effective on in-distribution (ID) data, GNNs often fail on out-of-distribution (OOD) nodes with unseen textual or structural patterns, producing overconfident predictions without reliable OOD detection. Existing topology-driven methods mitigate node-level bias through neighboring structures, but typically encode texts as shallow features, underutilizing semantic information. Recent LLM-based approaches instead synthesize pseudo OOD priors from textual knowledge, yet suffer from two key limitations: (1) a trade-off between reliability and informativeness, where generated OOD exposures either deviate from true OOD semantics or introduce substantial ID noise; and (2) dependence on specialized architectures, limiting compatibility with topology-level advances validated in prior work. To address these issues, we propose LG-Plug, an LLM-Guided Plug-and-play framework for TAG OOD detection. LG-Plug aligns topology and text representations to obtain fine-grained node embeddings, then constructs consensus-driven OOD exposure through clustered iterative LLM prompting. To reduce LLM query cost, it further adopts lightweight in-cluster codebooks and heuristic sampling. The generated OOD exposure acts as a regularizer that separates ID and OOD nodes, enabling seamless integration with existing detectors. Experiments on six TAG benchmarks demonstrate that LG-Plug consistently improves topology-driven OOD detectors (>7% FPR95 reduction) and surpasses prior LLM-based methods (>5% FPR95 reduction).

Figures

Figures reproduced from arXiv: 2602.11641 by Di Wu, Guocong Quan, Miao Hu, Xu Wang, Yinlin Zhu.

Figure 1
Figure 1. Figure 1: Illustration of the reliability-informativeness imbalance in LLM-based OOD prior generation: (a) Realistic ID and OOD categories [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed LG-Plug plug-and-play framework. (a) Topology-Text Representation Alignment jointly trains a graph encoder and a text encoder to learn fine-grained and discriminative node embeddings for both ID and OOD nodes. (b) Consensus-driven OOD Exposure with LLM identifies reliable and informative OOD exposures from unlabeled nodes by prompting a LLM while remaining scalable. The extracted O… view at source ↗
Figure 3
Figure 3. Figure 3: Distribution of negative energy scores on the Cora dataset. The dashed line indicates the threshold for a 95% true positive rate (TPR). [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis for the number of clusters (correspond [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Word cloud visualizations on the Cora dataset. (a)-(g) show ground-truth ID and OOD categories, while (h) displays all the OOD nodes. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

56 extracted references · 16 linked inside Pith

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Floren- cia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023)

    Pith/arXiv arXiv 2023

  2. [2]

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A next-generation hyperparameter optimization frame- work. InProceedings of the 25th ACM SIGKDD international conference on knowl- edge discovery & data mining, KDD. 2623–2631

    2019

  3. [3]

    Joan Bruna, Wojciech Zaremba, Arthur Szlam, and Yann LeCun. 2014. Spectral Networks and Locally Connected Networks on Graphs. arXiv:1312.6203 [cs.LG] https://arxiv.org/abs/1312.6203

    Pith/arXiv arXiv 2014

  4. [4]

    Tingyi Cai, Yunliang Jiang, Yixin Liu, Ming Li, Changqin Huang, and Shirui Pan. 2025. Out-of-Distribution Detection on Graphs: A Survey.arXiv preprint arXiv:2502.08105(2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Ziwei Chai, Tianjie Zhang, Liang Wu, Kaiqiang Han, Xiaohai Hu, Xuanwen Huang, and Yang Yang. 2025. Graphllm: Boosting graph reasoning ability of large language model.IEEE Transactions on Big Data(2025)

    2025

  6. [6]

    Zhikai Chen, Haitao Mao, Hongzhi Wen, Haoyu Han, Wei Jin, Haiyang Zhang, Hui Liu, and Jiliang Tang. 2023. Label-free node classification on graphs with large language models (llms).arXiv preprint arXiv:2310.04668(2023)

    Pith/arXiv arXiv 2023

  7. [7]

    Keyu Duan, Qian Liu, Tat-Seng Chua, Shuicheng Yan, Wei Tsang Ooi, Qizhe Xie, and Junxian He. 2023. Simteg: A frustratingly simple approach improves textual graph learning.arXiv preprint arXiv:2308.02565(2023)

    Pith/arXiv arXiv 2023

  8. [8]

    Jiarui Feng, Hao Liu, Lecheng Kong, Mingfang Zhu, Yixin Chen, and Muhan Zhang. 2024. TAGLAS: An atlas of text-attributed graph datasets in the era of large graph and language models. arXiv:2406.14683 [cs.LG] https://arxiv.org/ abs/2406.14683

    Pith/arXiv arXiv 2024

  9. [9]

    Shurui Gui, Xiner Li, Limei Wang, and Shuiwang Ji. 2022. Good: A graph out-of- distribution benchmark.Advances in Neural Information Processing Systems35 (2022), 2059–2073

    2022

  10. [10]

    Jiayan Guo, Lun Du, Hengyu Liu, Mengyu Zhou, Xinyi He, and Shi Han. 2023. Gpt4graph: Can large language models understand graph structured data? an empirical evaluation and benchmarking.arXiv preprint arXiv:2305.15066(2023)

    Pith/arXiv arXiv 2023

  11. [11]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs.Advances in neural information processing systems, NeurIPS(2017)

    2017

  12. [12]

    Yufei He and Bryan Hooi. 2024. Unigraph: Learning a cross-domain graph foundation model from natural language.CoRR(2024)

    2024

  13. [13]

    Dan Hendrycks and Kevin Gimpel. 2017. A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. In5th International Con- ference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net

    2017

  14. [14]

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, NeurIPS33 (2020), 22118–22133

    2020

  15. [15]

    Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations, ICLR

    2017

  16. [16]

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. Advances in neural information processing systems31 (2018)

    2018

  17. [17]

    Shiyu Liang, Yixuan Li, and R Srikant. 2018. Enhancing The Reliability of Out- of-distribution Image Detection in Neural Networks. InInternational Conference on Learning Representations

    2018

  18. [18]

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. 2023. One for all: Towards training one graph model for all classification tasks.arXiv preprint arXiv:2310.00149(2023)

    Pith/arXiv arXiv 2023

  19. [19]

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. 2020. Energy-based out-of-distribution detection.Advances in neural information processing systems 33 (2020), 21464–21475

    2020

  20. [20]

    Stuart Lloyd. 1982. Least squares quantization in PCM.IEEE transactions on information theory28, 2 (1982), 129–137

    1982

  21. [21]

    Xiangwei Lv, Mengze Li, Jingyuan Chen, Zhiang Dong, Sirui Han, and Beishui Liao. 2025. Out-of-Distribution Detection via LLM-Guided Outlier Generation for Text-attributed Graph. InFindings of the Association for Computational Linguistics: ACL 2025. 19544–19555

    2025

  22. [22]

    Longfei Ma, Yiyou Sun, Kaize Ding, Zemin Liu, and Fei Wu. 2024. Revisiting score propagation in graph out-of-distribution detection.Advances in Neural Information Processing Systems37 (2024), 4341–4373

    2024

  23. [23]

    Péter Mernyei and Cătălina Cangea. 2020. Wiki-cs: A wikipedia-based benchmark for graph neural networks.arXiv preprint arXiv:2007.02901(2020)

    Pith/arXiv arXiv 2020

  24. [24]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sand- hini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al

  25. [25]

    Xubin Ren, Jiabin Tang, Dawei Yin, Nitesh Chawla, and Chao Huang. 2024. A survey of large language models for graphs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6616–6626

    2024

  26. [26]

    Injy Sarhan and Marco Spruit. 2021. Open-cykg: An open cyber threat intelligence knowledge graph.Knowledge-based systems233 (2021), 107524

    2021

  27. [27]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective classification in network data.AI magazine29, 3 (2008), 93–93

    2008

  28. [28]

    Kartik Sharma, Yeon-Chang Lee, Sivagami Nambi, Aditya Salian, Shlok Shah, Sang-Wook Kim, and Srijan Kumar. 2024. A survey of graph neural networks for social recommender systems.Comput. Surveys56, 10 (2024), 1–34

    2024

  29. [29]

    Maximilian Stadler, Bertrand Charpentier, Simon Geisler, Daniel Zügner, and Stephan Günnemann. 2021. Graph posterior network: Bayesian predictive uncer- tainty for node classification.Advances in Neural Information Processing Systems 34 (2021), 18033–18048

    2021

  30. [30]

    Guangxin Su, Hanchen Wang, Jianwei Wang, Wenjie Zhang, Ying Zhang, and Jian Pei. 2025. Large Language Models Meet Text-Attributed Graphs: A Survey of Integration Frameworks and Applications.arXiv preprint arXiv:2510.21131 (2025)

    arXiv 2025

  31. [31]

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Lixin Su, Suqi Cheng, Dawei Yin, and Chao Huang. 2024. Graphgpt: Graph instruction tuning for large language models. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval. 491–500

    2024

  32. [32]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need.Advances in neural information processing systems30 (2017)

    2017

  33. [33]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. InInternational Con- ference on Learning Representations, ICLR

    2018

  34. [34]

    Danny Wang, Ruihong Qiu, Guangdong Bai, and Zi Huang. 2025. Gold: Graph out-of-distribution detection via implicit adversarial latent generation.arXiv preprint arXiv:2502.05780(2025)

    Pith/arXiv arXiv 2025

  35. [35]

    Heng Wang, Shangbin Feng, Tianxing He, Zhaoxuan Tan, Xiaochuang Han, and Yulia Tsvetkov. 2023. Can language models solve graph problems in natural language?Advances in Neural Information Processing Systems36 (2023), 30840– 30861

    2023

  36. [36]

    Zhihao Wen and Yuan Fang. 2023. Augmenting low-resource text classification with graph-grounded pre-training and prompting. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 506–516

    2023

  37. [37]

    Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. 2023. Energy- based out-of-distribution detection for graph neural networks.arXiv preprint arXiv:2302.02914(2023)

    Pith/arXiv arXiv 2023

  38. [38]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks.IEEE transactions on neural networks and learning systems32, 1 (2020), 4–24

    2020

  39. [39]

    Haoyan Xu, Zhengtao Yao, Ziyi Wang, Zhan Cheng, Xiyang Hu, Mengyuan Li, and Yue Zhao. 2025. Graph Synthetic Out-of-Distribution Exposure with Large Language Models. arXiv:2504.21198 [cs.LG] https://arxiv.org/abs/2504.21198

    Pith/arXiv arXiv 2025

  40. [41]

    Haoyan Xu, Zhengtao Yao, Xuzhi Zhang, Ziyi Wang, Langzhou He, Yushun Dong, Philip S Yu, Mengyuan Li, and Yue Zhao. 2025. GLIP-OOD: Zero-Shot Graph OOD Detection with Graph Foundation Model.arXiv preprint arXiv:2504.21186 (2025)

    Pith/arXiv arXiv 2025

  41. [42]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2019. How powerful are graph neural networks? (2019)

    2019

  42. [43]

    Hao Yan, Chaozhuo Li, Ruosong Long, Chao Yan, Jianan Zhao, Wenwen Zhuang, Jun Yin, Peiyan Zhang, Weihao Han, Hao Sun, et al . 2023. A Comprehensive Study on Text-attributed Graphs: Benchmarking and Rethinking.Advances in Neural Information Processing Systems36 (2023), 17238–17264

    2023

  43. [44]

    Hao Yan, Chaozhuo Li, Ruosong Long, Chao Yan, Jianan Zhao, Wenwen Zhuang, Jun Yin, Peiyan Zhang, Weihao Han, Hao Sun, Weiwei Deng, Qi Zhang, Lichao Sun, Xing Xie, and Senzhang Wang. 2023. A Comprehen- sive Study on Text-attributed Graphs: Benchmarking and Rethinking. InAd- vances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Glober- son,...

    2023

  44. [45]

    Shenzhi Yang, Bin Liang, An Liu, Lin Gui, Xingkai Yao, and Xiaofang Zhang. 2025. Bounded and uniform energy-based out-of-distribution detection for graphs. arXiv preprint arXiv:2504.13429(2025)

    Pith/arXiv arXiv 2025

  45. [46]

    Cohen, and Ruslan Salakhutdinov

    Zhilin Yang, William W. Cohen, and Ruslan Salakhutdinov. 2016. Revisiting Semi-Supervised Learning with Graph Embeddings. InProceedings of the 33rd International Conference on International Conference on Machine Learning, ICML. 40–48

    2016

  46. [47]

    Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang. 2024. Language is all a graph needs. InFindings of the association for computational linguistics: EACL 2024. 1955–1973

    2024

  47. [48]

    Jianxiang Yu, Yuxiang Ren, Chenghua Gong, Jiaqi Tan, Xiang Li, and Xuecang Zhang. 2025. Leveraging large language models for node generation in few-shot learning on text-attributed graphs. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 13087–13095

    2025

  48. [49]

    Taiyan Zhang, Renchi Yang, Yurui Lai, Mingyu Yan, Xiaochun Ye, and Dongrui Fan. 2025. Leveraging large language models for effective label-free node classi- fication in text-attributed graphs. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 698–708

    2025

  49. [50]

    Huanjing Zhao, Beining Yang, Yukuo Cen, Junyu Ren, Chenhui Zhang, Yuxiao Dong, Evgeny Kharlamov, Shu Zhao, and Jie Tang. 2024. Pre-training and prompt- ing for few-shot node classification on text-attributed graphs. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 4467–4478

    2024

  50. [51]

    Jianan Zhao, Meng Qu, Chaozhuo Li, Hao Yan, Qian Liu, Rui Li, Xing Xie, and Jian Tang. 2022. Learning on large-scale text-attributed graphs via variational inference.arXiv preprint arXiv:2210.14709(2022)

    Pith/arXiv arXiv 2022

  51. [52]

    Xujiang Zhao, Feng Chen, Shu Hu, and Jin-Hee Cho. 2020. Uncertainty aware semi-supervised learning on graph data.Advances in neural information process- ing systems33 (2020), 12827–12836

    2020

  52. [53]

    Yu Zhou, Haixia Zheng, Xin Huang, Shufeng Hao, Dengao Li, and Jumin Zhao

  53. [54]

    Yun Zhu, Haizhou Shi, Xiaotang Wang, Yongchao Liu, Yaoke Wang, Boci Peng, Chuntao Hong, and Siliang Tang. 2025. Graphclip: Enhancing transferability in graph foundation models for text-attributed graphs. InProceedings of the ACM on Web Conference 2025. 2183–2197

    2025

  54. [55]

    algorithm

    Yun Zhu, Yaoke Wang, Haizhou Shi, and Siliang Tang. 2024. Efficient tuning and inference for large language models on textual graphs.arXiv preprint arXiv:2401.15569(2024). Both Topology and Text Matter: Revisiting LLM-guided Out-of-Distribution Detection on Text-attributed Graphs A Dataset Descriptions and Processing. In this section, we introduce the dat...

    Pith/arXiv arXiv 2024

  55. [2021]

    In International conference on machine learning

    Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR, 8748–8763. Yinlin Zhu, Miao Hu, Di Wu

  56. [2022]

    Graph Neural Networks: Taxonomy, Advances, and Trends.ACM Transac- tions on Intelligent Systems and Technology (TIST)13, 1 (2022), 1–54

    2022