Pith. sign in

REVIEW 4 major objections 5 minor 73 references

GCoT: Chain-of-Thought Prompt Learning for Graphs

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

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

Pith's one-line read Adding iterative "thought" steps to graph prompts boosts few-shot accuracy

desk verdict A reasonable first stab at CoT-style prompting for text-free graphs; the empirical breadth is real, but the headline 'consistently outperforms' is not yet supported by the statistics or the ablations. read the letter →

arxiv 2502.08092 v2 pith:2ENDFVPL submitted 2025-02-12 cs.CL cs.AI

Xingtong Yu , Chang Zhou , Zhongwei Kuai , Xinming Zhang , Yuan Fang This is my paper · ORCID
classification cs.CLcs.AI
keywords graphpromptlearningchain-of-thoughtfew-shotclassificationnodecondition-netpre-trainedencoderiterativeinference
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 a frozen pre-trained graph encoder can be made to refine its predictions step by step, in analogy to chain-of-thought reasoning in language models, by interleaving prompt-based inference with "thought" construction and thought-conditioned prompt generation. The authors propose GCoT, a framework that repeatedly applies the same frozen encoder to multiplicatively reweighted node features, where the reweighting mask is generated from the encoder's own intermediate embeddings. Across one-shot node and graph classification on eight benchmark datasets, GCoT consistently outperforms fine-tuning, standard graph prompting, and pre-train-then-fine-tune baselines, and it also improves the accuracy of several existing graph prompting methods when plugged in as the standard prompt component.

What carries the argument

The load-bearing mechanism is the thought-conditioned prompt loop: at each inference step, the frozen graph encoder produces hidden layer embeddings, these are fused by learned weights into a thought matrix T_k, a lightweight MLP condition-net generates a node-specific prompt mask P_k from T_k, and element-wise multiplication P_k ⊙ X produces the input for the next pass. The thought acts as a compressed record of the model's current working state, and the condition-net converts that state into feature rescaling without adding a separate prompt vector per node.

What would settle it

Train GCoT and GCoT\CoT with identical numbers of trainable parameters and identical total forward passes (e.g., by widening the single-step prompt model), then compare one-shot accuracy: if the iterative version no longer beats the single-step model, the improvement is attributable to parameter count or inference budget rather than to thought-conditioned step-by-step reasoning.

Watch

Extended reading notes

Core claim

The central claim is that chain-of-thought-style iterative inference can be transferred to text-free graphs by replacing language thoughts with a fused representation of the frozen encoder's hidden layers, and by using a small conditional network to convert each node's thought into a node-specific multiplicative prompt. At step k the updated input is X_{k+1} = P_k ⊙ X with P_k = CondNet(Fuse($H^{1}$_k,...,H^L_k)), so the model refines its working state across K passes while the pre-trained weights stay frozen. The authors report that two steps for node classification and three steps for graph classification give the best results, and that the gap over the single-step variant GCoT\CoT confirms the value of iterative refinement rather than merely the added prompt parameters.

Load-bearing premise

The method's gains rest on the premise that repeatedly feeding the same frozen encoder a rescaling of the original features can genuinely refine predictions, an assumption the paper does not isolate from the added capacity of the condition-net and fusion weights.

Editorial extensions

If this is right

  • If GCoT is correct, existing single-step graph prompt methods can be upgraded by simply wrapping them in the iterative thought loop, gaining consistent few-shot accuracy improvements on both node and graph classification.
  • The step count becomes a tunable resource: node-level tasks saturate at K=2 while graph-level tasks benefit from K=3, and harder datasets like PROTEINS push the optimum to K=7, suggesting that task complexity controls the useful number of refinements.
  • The mechanism is compatible with both homophilic and heterophilic graphs, as shown by results on Wisconsin and Squirrel, extending its scope beyond the typical assortative benchmark setting.
  • Because the pre-trained encoder remains frozen and only the fusion weights, condition-net, and standard prompt are trained, the parameter overhead stays small even as the number of iterations grows, making the approach viable in few-shot regimes.

Reading between the lines

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

  • The multiplicative restriction X_{k+1} = P_k ⊙ X means the method can only reweight features the original graph already contains; it cannot synthesize new attributes. A natural testable extension would be to replace the product with an additive or gated combination and check whether the iterative gain persists, which would separate the benefit of step-wise refinement from the benefit of feature re
  • The analogy to language chain-of-thought is structural rather than semantic: the paper's "thoughts" are learned embedding mixtures, not interpretable reasoning steps. One could probe whether the thought T_k converges across steps or drifts, and correlate that with accuracy, which would clarify how closely the mechanism mimics genuine reasoning progress.
  • Since the standard prompt can be swapped among ProG, GPF, GPF+, and GraphPrompt with consistent gains, GCoT likely acts as an orthogonal refinement layer on top of any frozen-encoder prompt method, suggesting it could also be paired with contrastive graph foundation models that use a different task template.
  • A minimal control experiment would be to run GCoT with randomly initialized (untrained) thoughts and prompts: if accuracy still improves over GCoT\CoT, the gain would stem from the extra forward passes or the widened hypothesis space rather than from thought-conditioning specifically.
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. GCoT proposes a chain-of-thought style prompt-learning framework for frozen pre-trained graph encoders on text-free graphs. The method iterates K inference steps: at each step, a prompt-modified feature matrix is passed through a frozen encoder, the per-layer embeddings are fused into a "thought" T_k, and a condition-net maps T_k to node-specific prompts P_k, which update the input for the next step via X_{k+1} = P_k ⊙ X (Eq. 8). After K steps, a standard prompt (GPF+ style) modifies the output embedding before classification. Only the fusion weights, condition-net parameters, and standard prompts are trainable. Experiments cover one-shot and few-shot node/graph classification on eight datasets, heterophilic graphs, a large-scale CLUSTER test, ablations, hyperparameter sensitivity, and integration with several standard prompting methods. The authors claim consistent superiority over baselines in both node and graph classification.

Significance. If the central claim were established, GCoT would be a useful parameter-efficient mechanism for adapting frozen text-free graph encoders, and the paper's breadth—eight datasets, ablations, heterophily tests, a large-scale experiment, and a public code release—makes it a potentially valuable contribution to graph prompt learning. The core equations are simple and the method is reproducible in principle. However, the evidence as presented does not yet support the headline claims: most margins in Table 2 are within one standard deviation, the main ablation conflates the number of inference steps with the addition of the condition-net and fusion parameters, and the number of steps is selected in a way that is not clearly validation-based and is inconsistent for one dataset. The CoT framing is suggestive rather than demonstrated by the current experiments.

major comments (4)
  1. [Section 5.3, Table 2] The claim that "GCoT consistently outperforms the baseline methods across both node and graph classification" is not supported by the reported statistics because no significance tests or confidence intervals are provided. For example, on Cora GCoT scores 59.67 ± 15.51 versus 57.60 ± 13.88 for GPF and 57.42 ± 13.87 for GPF+; on Citeseer the margin is 46.21 ± 8.78 versus 45.34 ± 10.53 for GraphPrompt; and on PROTEINS the margin is 56.24 ± 8.60 versus 54.58 ± 8.70 for GPF+. With 100 sampled tasks and five seeds per task, paired statistical tests over the same task draws should be feasible and should be reported. Without such tests, phrases like "consistently outperforms" and "significant accuracy gains" (Section 5.8, Section 6) are not justified by the data.
  2. [Section 5.4, Table 3] The ablation GCoT\CoT does not isolate the effect of iterative inference. As described in the text, this variant "produces the final answer in a single inference step without CoT-style prompting," which removes not only the repeated encoder passes but also the entire thought-conditioned prompt path: the fusion weights in Eq. (6) and the condition-net in Eq. (7) are absent. Consequently, the accuracy gain of GCoT over GCoT\CoT could be due entirely to the additional trainable capacity of the condition-net and fusion, rather than to "step-by-step inference." A controlled comparison is needed, for example K=2 versus K=1 with the condition-net and fusion modules present and trained in both cases, or a variant that applies the generated prompt to the final embedding without re-encoding the graph. Without such an experiment, the conclusion that the advantage "underscores the importance of step-by-step inference" is not established.
  3. [Section 5.7, Fig. 6 and Section 5.2] The number of inference steps K is not handled in a way that supports the consistent-advantage claim. The main experiments fix K=2 for node classification and K=3 for graph classification, but Fig. 6 reports that PROTEINS peaks at K=7, meaning the main-result configuration is not the optimum for one of the eight datasets. The paper does not state whether K was chosen by validation or by inspecting test-set accuracy; if the latter, the reported results are post hoc. A validation-based selection procedure, or a report of results under a single fixed configuration, is required to make the comparison fair across datasets.
  4. [Section 4.2.1, Eqs. (7)-(8)] The mechanism is presented as a CoT-style refinement, but the recurrence is a deterministic map that only rescales the original features: X_{k+1} = CondNet(T_k) ⊙ X. Because the encoder is frozen and no new information is injected, the iteration cannot add features that were not already present in X; it can only reweight them. The paper provides no analysis of when this map refines predictions, whether it converges, or what role the number of steps K plays beyond increasing model capacity and computation. A brief contraction or sensitivity analysis, or at least an explicit acknowledgment of this restriction and an empirical check that the benefit is not merely added depth, would materially strengthen the central interpretation.
minor comments (5)
  1. [Section 5.5, Table 4] The text states that experiments are conducted on Wisconsin and Squirrel, but Table 4 is headed "Wisconsin Chameleon." The dataset description given in the text (5,201 nodes, 217,073 edges) matches Squirrel, not Chameleon; the table and the text should be made consistent.
  2. [Table 1 and Appendix A] Dataset statistics are inconsistent between Table 1 and Appendix A: Pubmed is listed with 88,648 edges in Table 1 but 44,338 citation links in Appendix A, and Photo is listed with 238,162 edges in Table 1 but 119,043 co-purchase edges in Appendix A. Please clarify whether one set uses directed edges and reconcile the numbers.
  3. [Section 5.4, last paragraph] The visualization paragraph says that two steps in total are used, but the caption of Fig. 4 and the surrounding text refer to "1st step thought" without stating the same K. Please state explicitly which configuration is visualized and whether the comparison is between GCoT\CoT and GCoT with the default K for the corresponding task.
  4. [Section 5.2] The condition-net hidden dimension is set to 32 for node classification and 8 for graph classification, while Fig. 5 suggests that the optimal s is dataset-dependent. The paper should state whether these values were selected on a validation split or are post hoc choices from the test results shown in Fig. 5.
  5. [Section 5.6, Table 5] For ProG, GPF, and GPF+, the table caption says these methods are applied to modify input features at the first inference step, but the main method description in Section 4.2.2 applies GPF+ to the output embeddings. The relationship between the standard prompt placement and the reported results should be clarified.
Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim is empirical and rests on standard pre-training assumptions plus two paper-specific design choices: the layer-fusion thought and multiplicative prompt application. No new physical or external entities are introduced; the 'thought' is an internal representation obtained by a weighted sum of existing hidden embeddings.

free parameters (2)
  • Number of inference steps K = 2 for node classification, 3 for graph classification; 7 reported best for PROTEINS in sensitivity analysis
    Selected based on accuracy across values in Fig. 6; directly controls how many times the encoder is applied, and hence the method's computation and accuracy.
  • Condition-net hidden dimension s = 32 for node classification, 8 for graph classification
    Selected via sensitivity analysis in Fig. 5; not derived from theory and influences expressiveness of generated prompts.
assumptions (4)
  • domain assumption The pre-trained graph encoder produces transferable representations for downstream node and graph classification.
    Section 4.1 assumes the frozen encoder (pre-trained with link prediction) provides a useful starting point for prompt-based adaptation.
  • ad hoc to paper A weighted sum of all hidden layers (the 'thought') captures a working state useful for generating prompts.
    Equation (6) in Section 4.2.1 defines T_k this way; the paper provides no analysis showing this fusion is superior to other aggregations beyond the ablation.
  • ad hoc to paper Element-wise multiplication of node features by the generated prompt is sufficient to guide the next inference step.
    Equation (8) in Section 4.2.1 restricts input modification to multiplicative masks of the original feature matrix; no additive or structural prompt is considered.
  • domain assumption Similarity-based classification with class prototypes, consistent with the pre-training objective, is an effective downstream head.
    Section 4.3, Eq. (11); follows the unified template from prior graph prompt methods [30, 59].

how reviews work

0 comments
Cite this review

Pith. "Pith review of GCoT: Chain-of-Thought Prompt Learning for Graphs." pith.science (2026). https://pith.science/paper/2ENDFVPL

@misc{pith2026250208092,
  author       = {Pith},
  title        = {Pith review of: GCoT: Chain-of-Thought Prompt Learning for Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2ENDFVPL}},
  note         = {Machine review of arXiv:2502.08092}
}
read the original abstract

Chain-of-thought (CoT) prompting has achieved remarkable success in natural language processing (NLP). However, its vast potential remains largely unexplored for graphs. This raises an interesting question: How can we design CoT prompting for graphs to guide graph models to learn step by step? On one hand, unlike natural languages, graphs are non-linear and characterized by complex topological structures. On the other hand, many graphs lack textual data, making it difficult to formulate language-based CoT prompting. In this work, we propose the first CoT prompt learning framework for text-free graphs, GCoT. Specifically, we decompose the adaptation process for each downstream task into a series of inference steps, with each step consisting of prompt-based inference, ``thought'' generation, and thought-conditioned prompt learning. While the steps mimic CoT prompting in NLP, the exact mechanism differs significantly. Specifically, at each step, an input graph, along with a prompt, is first fed into a pre-trained graph encoder for prompt-based inference. We then aggregate the hidden layers of the encoder to construct a ``thought'', which captures the working state of each node in the current step. Conditioned on this thought, we learn a prompt specific to each node based on the current state. These prompts are fed into the next inference step, repeating the cycle. To evaluate and analyze the effectiveness of GCoT, we conduct comprehensive experiments on eight public datasets, which demonstrate the advantage of our approach.

Figures

Figures reproduced from arXiv: 2502.08092 by the authors.

Figure 1
Figure 1. Illustration of chain-of-thought in NLP, standard [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of GCoT. 4.2.1 Inference step. Each inference step consists of three sub￾stages, as described in the following. Prompt-based inference. In the 𝑘-th inference step, we first feed the query graph𝐺 = (𝑉 , 𝐸), along with its prompt-modified feature matrix, into the pre-trained encoder: {H 1 𝑘 , H 2 𝑘 , · · · , H 𝐿 𝑘 } = GraphEncoder(X𝑘 ,𝐺; Θ0), (4) where Θ0 denotes the frozen pre-trained weights in the… view at source ↗
Figure 3
Figure 3. Impact of labeled data size (number of shots) on [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of node and thought embeddings. Dif [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 6
Figure 6. Figure 6: Impact of number of inference steps 𝐾 [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 5
Figure 5. Figure 5: Impact of hidden dimension [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 54 canonical work pages

  1. [1]

    Karsten M Borgwardt, Cheng Soon Ong, Stefan Schönauer, SVN Vishwanathan, Alex J Smola, and Hans-Peter Kriegel. 2005. Protein function prediction via graph kernels. Bioinformatics 21, suppl_1 (2005), i47–i56

    2005

  2. [2]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. NeurIPS 33 (2020), 1877–1901

    work page 2020

  3. [3]

    Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. 2023. A survey of chain of thought reasoning: Advances, frontiers and future.arXiv preprint arXiv:2309.15402 (2023)

    arXiv 2023

  4. [4]

    Asim Kumar Debnath, Rosa L Lopez de Compadre, Gargi Debnath, Alan J Shus- terman, and Corwin Hansch. 1991. Structure-activity relationship of mutagenic aromatic and heteroaromatic nitro compounds. correlation with molecular or- bital energies and hydrophobicity. Journal of medicinal chemistry 34, 2 (1991), 786–797

    1991

  5. [5]

    Vijay Prakash Dwivedi, Chaitanya K Joshi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. 2023. Benchmarking graph neural networks. Journal of Machine Learning Research 24, 43 (2023), 1–48

    work page 2023

  6. [6]

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. 2024. Universal prompt tuning for graph neural networks. NeurIPS (2024)

    work page 2024

  7. [7]

    Yuan Fang, Yuxia Wu, Xingtong Yu, and Shirui Pan. 2025. Few-Shot Learning on Graphs: From Meta-Learning to LLM-empowered Pre-Training and Beyond. In Companion Proceedings of WWW . 9–12

    work page 2025

  8. [8]

    Guhao Feng, Bohang Zhang, Yuntian Gu, Haotian Ye, Di He, and Liwei Wang

Show all 73 references
  1. [9]

    Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. 2023. Pal: Program-aided language models. In ICML. 10764–10799

    2023

  2. [10]

    David Ha, Andrew M Dai, and Quoc V Le. 2022. HyperNetworks. In ICLR

    2022

  3. [11]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. NeurIPS (2017), 1025–1035

    2017

  4. [12]

    Weihua Hu, Bowen Liu, Joseph Gomes, Marinka Zitnik, Percy Liang, Vijay Pande, and Jure Leskovec. 2020. Strategies for Pre-training Graph Neural Networks. In ICLR

    2020

  5. [13]

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, Kai-Wei Chang, and Yizhou Sun. 2020. GPT-GNN: Generative pre-training of graph neural networks. In SIGKDD. 1857– 1867

    2020

  6. [14]

    Shuo Ji, Xiaodong Lu, Mingzhe Liu, Leilei Sun, Chuanren Liu, Bowen Du, and Hui Xiong. 2023. Community-based dynamic graph learning for popularity prediction. In SIGKDD. 930–940

    2023

  7. [15]

    Xinke Jiang, Zidi Qin, Jiarong Xu, and Xiang Ao. 2023. Incomplete graph learning via attribute-structure decoupled variational auto-encoder. In WSDM. 304–312

    2023

  8. [16]

    Xinke Jiang, Rihong Qiu, Yongxin Xu, Wentao Zhang, Yichen Zhu, Ruizhe Zhang, Yuchen Fang, Xu Chu, Junfeng Zhao, and Yasha Wang. 2024. RAGraph: A General Retrieval-Augmented Graph Learning Framework. In NeurIPS

    2024

  9. [17]

    Xinke Jiang, Dingyi Zhuang, Xianghui Zhang, Hao Chen, Jiayuan Luo, and Xiaowei Gao. 2023. Uncertainty quantification via spatial-temporal tweedie model for zero-inflated and long-tail travel demand prediction. In CIKM

    2023

  10. [18]

    Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Zheng Li, Ruirui Li, Xianfeng Tang, Suhang Wang, Yu Meng, et al . 2024. Graph chain- of-thought: Augmenting large language models by reasoning on graphs. arXiv preprint arXiv:2404.07103 (2024)

    2024 arXiv

  11. [19]

    Anshul Kanakia, Zhihong Shen, Darrin Eide, and Kuansan Wang. 2019. A scalable hybrid research paper recommender system for microsoft academic. In WWW. 2893–2899

    2019

  12. [20]

    Thomas N Kipf and Max Welling. 2016. Variational graph auto-encoders. In Bayesian Deep Learning Workshop

    2016

  13. [21]

    Thomas N Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In ICLR

    2017

  14. [22]

    Namkyeong Lee, Kanghoon Yoon, Gyoung S Na, Sein Kim, and Chanyoung Park

  15. [23]

    Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The Power of Scale for Parameter-Efficient Prompt Tuning. In EMNLP. 3045–3059

    2021

  16. [24]

    Yibo Li, Yuan Fang, Mengmei Zhang, and Chuan Shi. 2024. FineMolTex: Towards Fine-grained Molecular Graph-Text Pre-training. arXiv preprint arXiv:2409.14106 (2024)

    2024 arXiv

  17. [25]

    Yibo Li, Xiao Wang, Hongrui Liu, and Chuan Shi. 2024. A Generalized Neural Diffusion Framework on Graphs. In AAAI. 8707–8715

    2024

  18. [26]

    Yibo Li, Xiao Wang, Yujie Xing, Shaohua Fan, Ruijia Wang, Yaoqi Liu, and Chuan Shi. 2024. Graph Fairness Learning under Distribution Shifts. InWWW. 676–684

    2024

  19. [27]

    Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021. GPT understands, too. arXiv preprint arXiv:2103.10385 (2021)

    2021 arXiv

  20. [28]

    Yang Liu, Deyu Bo, Wenxuan Cao, Yuan Fang, Yawen Li, and Chuan Shi. 2025. Graph Positional Autoencoders as Self-supervised Learners. arXiv:2505.23345

    2025 arXiv

  21. [29]

    Zemin Liu, Yuan Fang, Chenghao Liu, and Steven CH Hoi. 2021. Relative and absolute location embedding for few-shot node classification on graph. In AAAI. 4267–4275

    2021

  22. [30]

    Zemin Liu, Xingtong Yu, Yuan Fang, and Xinming Zhang. 2023. GraphPrompt: Unifying pre-training and downstream tasks for graph neural networks. InWWW. 417–428

    2023

  23. [31]

    Yuanfu Lu, Xunqiang Jiang, Yuan Fang, and Chuan Shi. 2021. Learning to pre-train graph neural networks. In AAAI. 4276–4284

    2021

  24. [32]

    Andrew Kachites McCallum, Kamal Nigam, Jason Rennie, and Kristie Seymore

  25. [33]

    Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang

  26. [34]

    Jiezhong Qiu, Qibin Chen, Yuxiao Dong, Jing Zhang, Hongxia Yang, Ming Ding, Kuansan Wang, and Jie Tang. 2020. GCC: Graph contrastive coding for graph neural network pre-training. In SIGKDD. 1150–1160

    2020

  27. [35]

    Rossi and Nesreen K

    Ryan A. Rossi and Nesreen K. Ahmed. 2015. The Network Data Repository with Interactive Graph Analytics and Visualization. In AAAI. 4292–4293

    2015

  28. [36]

    Benedek Rozemberczki, Carl Allen, and Rik Sarkar. 2021. Multi-scale attributed node embedding. Journal of Complex Networks (2021), cnab014

    2021

  29. [37]

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

    2008

  30. [38]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868 (2018)

    2018 arXiv

  31. [39]

    Fan-Yun Sun, Jordan Hoffmann, Vikas Verma, and Jian Tang. 2020. Infograph: Un- supervised and semi-supervised graph-level representation learning via mutual information maximization. In ICLR

    2020

  32. [40]

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. 2022. GPPT: Graph Pre-training and Prompt Tuning to Generalize Graph Neural Networks. In SIGKDD. 1717–1727

    2022

  33. [41]

    Xiangguo Sun, Hong Cheng, Jia Li, Bo Liu, and Jihong Guan. 2023. All in one: Multi-task prompting for graph neural networks. In SIGKDD. 2120–2131

    2023

  34. [42]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. 2018. Graph attention networks. In ICLR

    2018

  35. [43]

    Petar Velickovic, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. 2019. Deep Graph Infomax. In ICLR

    2019

  36. [44]

    Ning Wang, Minnan Luo, Kaize Ding, Lingling Zhang, Jundong Li, and Qinghua Zheng. 2020. Graph few-shot learning with attribute matching. In CIKM. 1545– 1554

    2020

  37. [45]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. ICLR (2023)

    2023

  38. [46]

    Xu Wang, Huan Zhao, Wei-wei Tu, and Quanming Yao. 2023. Automated 3D pre-training for molecular property prediction. In SIGKDD. 2419–2430

    2023

  39. [47]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. NeurIPS 35 (2022), 24824–24837

    2022

  40. [48]

    Zhihao Wen and Yuan Fang. 2023. Augmenting Low-Resource Text Classification with Graph-Grounded Pre-training and Prompting. In SIGIR. 506–516

    2023

  41. [49]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. TNNLS 32, 1 (2020), 4–24

    2020

  42. [50]

    Chenyan Xiong, Russell Power, and Jamie Callan. 2017. Explicit semantic ranking for academic search via knowledge graph embedding. In WWW. 1271–1279

    2017

  43. [51]

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

    2019

  44. [52]

    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.NeurIPS 36 (2023), 17238–17264

    2023

  45. [53]

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. 2024. Tree of thoughts: Deliberate problem solving with large language models. NeurIPS 36 (2024)

    2024

  46. [54]

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. 2021. Do transformers really perform badly for graph representation?. In NeurIPS. 28877–28888

    2021

  47. [55]

    Yuning You, Tianlong Chen, Yongduo Sui, Ting Chen, Zhangyang Wang, and Yang Shen. 2020. Graph contrastive learning with augmentations. NeurIPS 33 (2020), 5812–5823

    2020

  48. [56]

    Xingtong Yu, Yuan Fang, Zemin Liu, Yuxia Wu, Zhihao Wen, Jianyuan Bo, Xin- ming Zhang, and Steven CH Hoi. 2024. Few-Shot Learning on Graphs: from Meta-learning to Pre-training and Prompting. arXiv preprint arXiv:2402.01440 (2024). GCoT: Chain-of-Thought Prompt Learning for Gra...

    2024 arXiv

  49. [57]

    Xingtong Yu, Yuan Fang, Zemin Liu, and Xinming Zhang. 2024. Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning. In AAAI, Vol. 38. 16578–16586

    2024

  50. [58]

    Xingtong Yu, Zechuan Gong, Chang Zhou, Yuan Fang, and Hui Zhang. 2025. SAMGPT: Text-free graph foundation model for multi-domain pre-training and cross-domain adaptation. In Proceedings of the ACM on Web Conference 2025 . 1142–1153

    2025

  51. [59]

    Xingtong Yu, Zhenghao Liu, Yuan Fang, Zemin Liu, Sihong Chen, and Xinming Zhang. 2024. Generalized graph prompt: Toward a unification of pre-training and downstream tasks on graphs. IEEE TKDE (2024)

    2024

  52. [60]

    Xingtong Yu, Zemin Liu, Yuan Fang, and Xinming Zhang. 2023. Learning to count isomorphisms with graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 4845–4853

    2023

  53. [61]

    Xingtong Yu, Zhenghao Liu, Xinming Zhang, and Yuan Fang. 2025. Node-Time Conditional Prompt Learning In Dynamic Graphs. In ICLR

    2025

  54. [62]

    Xingtong Yu, Jie Zhang, Yuan Fang, and Renhe Jiang. 2025. Non-homophilic graph pre-training and prompt learning. In SIGKDD

    2025

  55. [63]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. Multigprompt for multi-task pre-training and prompting on graphs. In WWW. 515–526

    2024

  56. [64]

    Xingtong Yu, Chang Zhou, Yuan Fang, and Xinming Zhang. 2024. Text-Free Multi-domain Graph Pre-training: Toward Graph Foundation Models. arXiv preprint arXiv:2405.13934 (2024)

    2024 arXiv

  57. [65]

    Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim

  58. [66]

    Shiqi Zhang, Yiqian Huang, Jiachen Sun, Wenqing Lin, Xiaokui Xiao, and Bo Tang. 2023. Capacity constrained influence maximization in social networks. In SIGKDD. 3376–3385

    2023

  59. [67]

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications. AI open (2020), 57–81

    2020

  60. [68]

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. 2022. Conditional prompt learning for vision-language models. In CVPR. 16816–16825. Appendices A Further Descriptions of Datasets We provide more comprehensive descriptions of the benchmark datasets345 used in our ex...

    2022

  61. [2000]

    Information Retrieval (2000)

    Automating the construction of internet portals with machine learning. Information Retrieval (2000)

    2000

  62. [2019]

    NeurIPS 32 (2019)

    Graph transformer networks. NeurIPS 32 (2019)

    2019

  63. [2020]

    arXiv preprint arXiv:2002.05287 (2020)

    Geom-gcn: Geometric graph convolutional networks. arXiv preprint arXiv:2002.05287 (2020)

    2020 arXiv

  64. [2023]

    In SIGKDD

    Shift-robust molecular relational learning with causal substructure. In SIGKDD. 1200–1212

  65. [2024]

    NeurIPS 36 (2024)

    Towards revealing the mystery behind chain of thought: a theoretical perspective. NeurIPS 36 (2024)

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.