Pith. sign in

REVIEW 3 major objections 4 minor 76 references

Quantizing Text-attributed Graphs for Semantic-Structural Integration

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read Quantized graph tokens let frozen LLMs classify nodes with zero labels.

desk verdict Novel quantized-token framework for text-attributed graphs, but the zero-shot transfer claim is overstated and the evaluation needs cleanup before it can be trusted. read the letter →

arxiv 2507.19526 v1 pith:3AY47LPX submitted 2025-07-20 cs.LG cs.AI

classification cs.LGcs.AI
keywords text-attributedgraphslargelanguagemodelsgraphquantizationzero-shottransferfew-shotlearningself-supervisednodeclassificationsofttokenization
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes STAG (Soft Tokenization for Text-attributed Graphs), a self-supervised framework that turns each node of a text-attributed graph—a graph whose nodes carry textual descriptions—into a short list of discrete vocabulary tokens, so a frozen large language model (LLM) can classify the node directly. The central claim is that structural information can be folded into these tokens without learned projectors or manual graph verbalization, by softly quantizing fused semantic and structural embeddings against a fixed codebook. If the claim is right, cross-dataset transfer no longer needs labels even in the source domain, and one pretrained graph model can serve many different LLMs. The paper supports this with experiments on few-shot and zero-shot node classification, plus link prediction, edge classification, and subgraph classification.

What carries the argument

STAG's central machinery is soft tokenization with a frozen semantic codebook. The codebook is built by embedding 15,062 filtered subword tokens from LLaMA-2 with the same frozen sentence transformer that embeds node text, so every codeword remains an ordinary word an LLM can tokenize. Instead of hard nearest-neighbor lookup, Eq. (2) computes temperature-softmax cosine attention over the full codebook, Eq. (3) forms the quantized vector as the weighted sum, and Eq. (5) adds a KL divergence that pulls the fused-feature attention distribution toward the original text attention distribution. This is the mechanism that carries structural and semantic information into discrete tokens while keeping them interpretable by any LLM sharing the vocabulary.

What would settle it

Replace the adjacency matrix with the identity matrix during STAG pre-training, keeping all losses and the codebook unchanged, then measure zero-shot classification on a cross-domain target; if accuracy does not drop, structural information is not actually entering the tokens.

Watch

Extended reading notes

Core claim

The paper's central claim is that graph structure can be made legible to an LLM without training an alignment layer. During pre-training, a graph neural network encoder produces structural node embeddings; these are fused with sentence-transformer text features, and the fused vector is softly assigned over a frozen codebook of 15,062 embeddings of LLaMA-2 subword tokens using temperature-scaled cosine attention. A KL-divergence loss aligns this assignment with the assignment of the original text, keeping it semantically meaningful, while reconstruction and contrastive losses preserve node-level semantics and neighborhood structure. At inference, the top tokens are used directly as the node prompt for an LLM, or the fused features feed a linear probe or prompt-tuned classifier. The paper claims this achieves true zero-shot cross-dataset transfer, because pre-training is label-free, and that the same tokens work across LLaMA-2, LLaMA-3, Vicuna, and GPT-4o.

Load-bearing premise

The method assumes that cosine similarity in the frozen sentence-transformer embedding space is a faithful measure of semantic relatedness, so that the KL-aligned soft assignment selects tokens whose meaning an arbitrary LLM will recognize; if that geometric similarity does not track semantic relevance, the quantized tokens are arbitrary and the LLM path adds nothing over raw text.

Editorial extensions

If this is right

  • A single pre-trained STAG model can be paired with different LLMs at inference time without retraining, since the tokens come from a shared subword vocabulary.
  • Cross-dataset transfer no longer requires labeled source data, because pre-training is self-supervised and the same frozen model can be evaluated zero-shot on an unseen dataset.
  • GraphLLM pipelines can drop per-model projector networks and manual graph verbalization; the LLM is used frozen and consumes only the quantized token list plus a system prompt.
  • A non-LLM classifier on the fused features provides a cheap fallback path, so the framework does not depend on API availability for deployment.

Reading between the lines

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

  • Because the codebook is frozen and built from ordinary subword tokens, the same recipe should transfer to TAGs in other languages or domains by swapping the vocabulary, without retraining the graph encoder.
  • The class-specific codebook used in prompt tuning implies a general pattern: any task whose classes can be described in text can be reduced to a fixed embedding codebook, so STAG-style quantization could extend beyond node classification to relation prediction or retrieval.
  • A direct test of whether structure actually reaches the tokens would be to pre-train STAG with the adjacency matrix replaced by the identity matrix; if zero-shot accuracy on cross-domain targets does not drop, then the structural branch is not doing the work.
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 / 4 minor

Summary. The paper proposes STAG, a self-supervised framework that quantizes text-attributed graph (TAG) nodes into discrete tokens from a frozen codebook for use by large language models (LLMs) or by traditional classifiers. Node text is embedded with a frozen sentence transformer, a codebook is built from LLaMA-2 subword tokens embedded with the same transformer, and a GNN encoder is pre-trained using a fusion of structural and semantic features, soft assignment to codebook tokens, a KL-divergence alignment loss, reconstruction, and contrastive objectives. At inference, STAG can prompt an LLM with the top-k codebook tokens per node, or can be used without an LLM via linear probing or a prompt-tuning module. The main claims are that this enables 'true zero-shot transfer learning' without any source or target labels, that STAG 'consistently outperforms' raw-feature and raw-text baselines, and that it achieves state-of-the-art results on several node classification benchmarks.

Significance. If the central claim is correct, STAG would offer a practical and elegant solution to a well-known difficulty in GraphLLM research: embedding structural information into discrete LLM-compatible tokens without learned projection layers or manual graph verbalization, and without requiring labeled data in the source domain. The paper is empirically extensive, covering seven datasets, multiple LLM backbones, ablations, and auxiliary tasks (link prediction, edge classification, subgraph classification). The design is plausible: the codebook is frozen, the loss functions are self-supervised, and the ablation study shows that each component contributes to the reported performance. However, the load-bearing zero-shot transfer claim is weakened by the inclusion of in-domain target datasets in the main zero-shot table and by held-out results that contradict the 'consistently outperforms' assertion. These issues are central to the paper's stated contribution and need to be addressed before the claims can be accepted.

major comments (3)
  1. [Section 4.3, Tables 2 and 10] The zero-shot transfer evaluation is partly in-domain. In Table 2, the model is pre-trained on Cora Full and is then evaluated on Cora and Cora Full as target datasets; these numbers measure in-domain performance, not transfer. The genuinely held-out datasets in Table 10 show the opposite of the paper's claim that STAG 'consistently outperforms' raw-feature variants: STAG + C loses to Raw Feat + C on PubMed (61.80 vs 63.50) and ogbn-products (73.17 vs 74.66), and the LLM path loses to Raw Feat + Q on PubMed (34.85 vs 42.35) and ogbn-products (61.58 vs 62.42). The claim of consistent, label-free zero-shot transfer is therefore not supported by the paper's own held-out numbers. I would recommend reporting zero-shot results only on datasets not used in pre-training, and softening or removing the 'consistently outperforms' wording unless the discrepancy is resolved.
  2. [Section 4.1, top-k selection] The manuscript states that 'For LLM inference, we use top-k = 13 tokens based on empirical performance' but does not specify the selection procedure, the validation split used, or whether target datasets were consulted during this selection. If top-k was tuned on each target dataset or on a validation subset of the target, the reported zero-shot and few-shot numbers are not fully label-free. Please clarify the hyperparameter selection protocol and, if possible, report sensitivity to top-k.
  3. [Section 3.2, Eq. (5), and Section 4.5] The method's core assumption is that cosine similarity in the frozen sentence-transformer embedding space, used in Eq. (2), produces tokens that an arbitrary LLM can interpret as semantically meaningful node descriptions. The KL alignment in Eq. (5) ensures that the fused representation attends similarly to the codebook as the text-only representation, but it does not by itself guarantee that the LLM's token-level semantics match the sentence-transformer geometry. The qualitative examples in Section 4.5 are anecdotal and cherry-picked; they do not quantify how often the quantized tokens are semantically coherent or how much the LLM path relies on the structural versus semantic components. Given that Table 10 shows the LLM path can lose to raw-feature quantization on held-out datasets, I would like to see a systematic evaluation of token semantics (e.g., token overlap with gold labels, human or LLM-judged coherence, or a diagnostic that varies structure while holding text fixed).
minor comments (4)
  1. [Table 10 and Section 4.3] The naming of the raw-feature baselines is inconsistent: Table 2 and Table 10 use 'Raw Feat + Q' and 'Raw Feat + C', while Table 1 uses 'Raw Feat + Quantization' and 'Raw Feat + Linear Probing'. Please unify the notation.
  2. [Equation (8)] The subscripts in 'L Rec commit' and 'L Contrast commit' are awkward and could be confused with products; please use a cleaner notation such as L_commit^Rec and L_commit^Con.
  3. [Section 4.2, Table 1] The caption says 'Gray-shaded rows: supervised baselines trained directly on target datasets', but the rows are not shaded in the text version, making it hard to identify them. Consider adding actual shading or explicit row markers in the printed table.
  4. [Section 4.2] The phrase 'despite raw texts' potential information leakage' is unclear; raw text is the input signal, not a leak. Rephrase to clarify what is meant.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: STAG's token outputs are functions of unlabeled text and graph structure, with no fitted parameter renamed as prediction.

full rationale

STAG's derivation chain is self-contained. Initial node features X are produced by a frozen sentence transformer applied to raw text (Section 3.2), and the codebook E is built by embedding filtered LLaMA-2 vocabulary tokens with the same frozen transformer. The GNN encoder e(·) is trained on unlabeled graphs with reconstruction (Eq. 6), contrastive (Eq. 7), commitment (Eq. 4), and KL-alignment (Eq. 5) losses; none of these objectives uses target labels or a fitted parameter that is later renamed as a prediction. At inference, tokens are selected by top-k attention over the frozen codebook (Eq. 9), so the output is a deterministic function of text, adjacency, and the self-supervised parameters. The KL term does align attn(z_f) with attn(x), so in the limit of full alignment the token distribution would match the text-only attention distribution, but the final objective (Eq. 8) also contains structural contrastive and reconstruction terms, so structural information can shift z_f; this is a regularizer, not a definitional identity. Self-citations ([61], [37]) are used for component-level methodology (prompt tuning and a subgraph-classification protocol), not to justify the central quantization claim. The inclusion of Cora Full and Cora as zero-shot targets after pre-training on Cora Full is an evaluation-design concern rather than a circular derivation, and held-out results are also reported (Tables 2 and 10). No predicted quantity reduces by construction to an input quantity.

Assumptions & free parameters 8 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are postulated; the codebook is built from existing LLaMA-2 vocabulary and sentence-transformer embeddings, and the class-specific codebook is a prompt-tuning artifact. The central method depends on many tuned hyperparameters, but no target quantity is derived from a fitted parameter, so this is an empirical method paper rather than a derivation.

free parameters (8)
  • phi, psi (feature fusion weights) = learned during pretraining, not reported
    Eq. (1) uses these weights to balance structural and semantic features; the claimed structure-aware tokens depend directly on them.
  • tau_sa (soft assignment temperature) = not reported
    Eq. (2) uses tau_sa to control softness of codebook attention; it is absent from the hyperparameter table, leaving reimplementation incomplete.
  • beta (commitment strength) = 1.9, 0.58, 1.4 per pretraining dataset
    Eq. (4) commitment loss strength, tuned per dataset with Optuna and varied across Cora Full, ogbn-arxiv, and ogbn-products.
  • lambda (KL alignment weight) = 1.0, 1.0, 1.6 per pretraining dataset
    Eq. (8) balances the KL alignment loss and therefore controls how strongly quantized tokens stay close to text-only semantics.
  • tau_c (contrastive temperature) = 0.831, 0.354, 0.103 per pretraining dataset
    Eq. (7) temperature for the neighborhood contrastive loss, tuned per dataset.
  • top-k (number of tokens per node) = 13 for LLM inference, 5 for closed-source LLMs
    Selected 'based on empirical performance' in Section 4.1; no sensitivity analysis is reported, so the LLM inference claim depends on this choice.
  • mask rate = 0.53, 0.6, 0.74 per pretraining dataset
    GraphMAE-style masking fraction used in the contrastive branch, affecting the structural signal available during pretraining.
  • optimization hyperparameters (learning rate, weight decay, hidden dim, layers, heads) = listed per dataset in Table 8
    Tuned separately for each pretraining dataset with Optuna, but the validation protocol for choosing them is not described.
assumptions (3)
  • domain assumption Cosine similarity in sentence-transformer embedding space is semantically meaningful for selecting codebook tokens.
    Enters at Eq. (2); if sentence-transformer cosine geometry does not align with semantic compatibility, soft assignment selects arbitrary tokens and the LLM inference path degenerates.
  • domain assumption Masked-node contrastive pairs of decoded and original features capture neighborhood structure.
    Section 3.3.3 assumes that treating other masked nodes' decoded features as negatives pushes the quantized representation to encode local graph structure rather than only node text.
  • domain assumption Filtered LLaMA-2 subwords embedded by a sentence transformer remain interpretable by other LLMs such as LLaMA3, Vicuna, and GPT-4o.
    Section 3.2 and Table 3 claim cross-architecture compatibility because tokens come from one LLM vocabulary, yet the embedding geometry is produced by a different model family.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Quantizing Text-attributed Graphs for Semantic-Structural Integration." pith.science (2026). https://pith.science/paper/3AY47LPX

@misc{pith2026250719526,
  author       = {Pith},
  title        = {Pith review of: Quantizing Text-attributed Graphs for Semantic-Structural Integration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3AY47LPX}},
  note         = {Machine review of arXiv:2507.19526}
}
read the original abstract

Text-attributed graphs (TAGs) have emerged as a powerful representation for modeling complex relationships across diverse domains. With the rise of large language models (LLMs), there is growing interest in leveraging their capabilities for graph learning. However, current approaches face significant challenges in embedding structural information into LLM-compatible formats, requiring either computationally expensive alignment mechanisms or manual graph verbalization techniques that often lose critical structural details. Moreover, these methods typically require labeled data from source domains for effective transfer learning, significantly constraining their adaptability. We propose STAG, a novel self-supervised framework that directly quantizes graph structural information into discrete tokens using a frozen codebook. Unlike traditional quantization approaches, our method employs soft assignment and KL divergence guided quantization to address the unique challenges of graph data, which lacks natural tokenization structures. Our framework enables both LLM-based and traditional learning approaches, supporting true zero-shot transfer learning without requiring labeled data even in the source domain. Extensive experiments demonstrate state-of-the-art performance across multiple node classification benchmarks while maintaining compatibility with different LLM architectures, offering an elegant solution to bridging graph learning with LLMs.

Figures

Figures reproduced from arXiv: 2507.19526 by the authors.

Figure 1
Figure 1. Overview of STAG framework: (a) Feature extraction and codebook construction, (b) Self-supervised pre-training with [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 58 canonical work pages

  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)

  2. [2]

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna: A Next-generation Hyperparameter Optimization Frame- work. In Proceedings of the 25th ACM SIGKDD International Conference on Knowl- edge Discovery and Data Mining

  3. [3]

    Antoine Bordes, Nicolas Usunier, Alberto García-Durán, Jason Weston, and Ok- sana Yakhnenko. 2013. Translating Embeddings for Modeling Multi-relational Data. In Advances in Neural Information Processing Systems , Christopher J. C. Burges, Léon Bottou, Zoubin Ghahramani, and Kilian Q. Weinberger (Eds.)

  4. [4]

    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. Advances in neural information processing systems (2020)

  5. [5]

    Runjin Chen, Tong Zhao, Ajay Kumar Jaiswal, Neil Shah, and Zhangyang Wang

  6. [6]

    Zhikai Chen, Haitao Mao, Hang Li, Wei Jin, Hongzhi Wen, Xiaochi Wei, Shuaiqiang Wang, Dawei Yin, Wenqi Fan, Hui Liu, et al . 2024. Exploring the potential of large language models (llms) in learning on graphs. ACM SIGKDD Explorations Newsletter (2024)

  7. [7]

    Gonzalez, Ion Stoica, and Eric P

    Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. 2023. Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality. https://lmsys.org/blog/2023-03-30-vicuna/

  8. [8]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Ima- genet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

Show all 76 references
  1. [9]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 (2024)

  2. [10]

    Patrick Esser, Robin Rombach, and Bjorn Ommer. 2021. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

  3. [11]

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin

  4. [12]

    Bahare Fatemi, Jonathan Halcrow, and Bryan Perozzi. 2024. Talk like a Graph: Encoding Graphs for Large Language Models. In International Conference on Learning Representations

  5. [13]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining

  6. [14]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al . 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948 (2025)

  7. [15]

    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)

  8. [16]

    Kaveh Hassani and Amir Hosein Khasahmadi. 2020. Contrastive multi-view rep- resentation learning on graphs. In International Conference on Machine Learning

  9. [17]

    Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2021. Deberta: decoding-Enhanced Bert with Disentangled Attention. InInternational Conference on Learning Representations

  10. [18]

    Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, and Bryan Hooi. 2024. Harnessing Explanations: LLM-to-LM Interpreter for Enhanced Text-Attributed Graph Representation Learning. In International Conference on Learning Representations

  11. [19]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. In Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval

  12. [20]

    Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi

    Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V. Chawla, Thomas Laurent, Yann LeCun, Xavier Bresson, and Bryan Hooi. 2024. G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering. In Ad- vances in Neural Information Processing Systems

  13. [21]

    Zhongmou He, Jing Zhu, Shengyi Qian, Joyce Chai, and Danai Koutra. 2024. LinkGPT: Teaching Large Language Models To Predict Missing Links. arXiv preprint arXiv:2406.04640 (2024)

  14. [22]

    Jordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, Elena Buchatskaya, Trevor Cai, Eliza Rutherford, Diego de Las Casas, Lisa Anne Hendricks, Johannes Welbl, Aidan Clark, et al. 2022. Training compute-optimal large language models. arXiv preprint arXiv:2203.15556 (2022)

  15. [23]

    Zhenyu Hou, Yufei He, Yukuo Cen, Xiao Liu, Yuxiao Dong, Evgeny Kharlamov, and Jie Tang. 2023. Graphmae2: A decoding-enhanced masked self-supervised graph learner. In Proceedings of the ACM web conference

  16. [24]

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. 2022. Graphmae: Self-supervised masked graph autoencoders. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

  17. [25]

    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. In Advances in neural information processing systems

  18. [26]

    Qian Huang, Hongyu Ren, Peng Chen, Gregor Kržmanc, Daniel Zeng, Percy S Liang, and Jure Leskovec. 2024. Prodigy: Enabling in-context learning over graphs. In Advances in Neural Information Processing Systems

  19. [27]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276 (2024)

  20. [28]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical Reparameterization with Gumbel-Softmax. In International Conference on Learning Representations

  21. [29]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of NaacL-HLT

  22. [30]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Repre- sentations

  23. [31]

    Thomas N Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Repre- sentations

  24. [32]

    Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. 2022. Autoregressive image generation using residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  25. [33]

    Yuhan Li, Peisong Wang, Zhixun Li, Jeffrey Xu Yu, and Jia Li. 2024. Zerog: Investigating cross-dataset zero-shot transferability in graphs. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

  26. [34]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. In Computer Vision–ECCV 2014: 13th European Conference

  27. [35]

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. 2024. One For All: Towards Training One Graph Model For All Classification Tasks. In International Conference on Learning Representations

  28. [36]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled Weight Decay Regularization. In International Conference on Learning Representations

  29. [37]

    Yuanfu Lu, Xunqiang Jiang, Yuan Fang, and Chuan Shi. 2021. Learning to pre- train graph neural networks. In Proceedings of the AAAI conference on artificial intelligence

  30. [38]

    Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel

  31. [39]

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

  32. [40]

    Péter Mernyei and Cătălina Cangea. 2020. Wiki-CS: A Wikipedia-Based Bench- mark for Graph Neural Networks. arXiv preprint arXiv:2007.02901 (2020)

  33. [41]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

  34. [42]

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. Deepwalk: Online learning of social representations. In Proceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining

  35. [43]

    Haoxuan Qu, Yujun Cai, and Jun Liu. 2024. Llms are good action recogniz- ers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

  36. [44]

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. 2021. Zero-shot text-to-image generation. In International conference on machine learning

  37. [45]

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. 2019. Generating diverse high-fidelity images with vq-vae-2. In Advances in neural information processing systems

  38. [46]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. In Conference on Empirical Methods in Natural Language Processing

  39. [47]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Quantizing Text-attributed Graphs for Semantic-Structural Integration KDD ’25, August 3–7, 2025, Toronto, ON, Canada Proceed...

  40. [48]

    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 Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

  41. [49]

    Xiangguo Sun, Hong Cheng, Jia Li, Bo Liu, and Jihong Guan. 2023. All in one: Multi-task prompting for graph neural networks. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining

  42. [50]

    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. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval

  43. [51]

    Shantanu Thakoor, Corentin Tallec, Mohammad Gheshlaghi Azar, Mehdi Azabou, Eva L Dyer, Remi Munos, Petar Veličković, and Michal Valko. 2022. Large-scale representation learning on graphs via bootstrapping. In International Conference on Learning Representations

  44. [52]

    Yijun Tian, Huan Song, Zichen Wang, Haozhu Wang, Ziqing Hu, Fang Wang, Nitesh V Chawla, and Panpan Xu. 2024. Graph neural prompting with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence

  45. [53]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yas- mine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhos- ale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288 (2023)

  46. [54]

    Théo Trouillon, Johannes Welbl, Sebastian Riedel, Éric Gaussier, and Guillaume Bouchard. 2016. Complex embeddings for simple link prediction. In International conference on machine learning

  47. [55]

    Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Neural Discrete Representation Learning. In Advances in Neural Information Processing Systems

  48. [56]

    Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. In International Con- ference on Learning Representations

  49. [57]

    Hamilton, Pietro Liò, Yoshua Bengio, and R

    Petar Veličković, William Fedus, William L. Hamilton, Pietro Liò, Yoshua Bengio, and R. Devon Hjelm. 2019. Deep Graph Infomax. In International Conference on Learning Representations

  50. [58]

    Heng Wang, Shangbin Feng, Tianxing He, Zhaoxuan Tan, Xiaochuang Han, and Yulia Tsvetkov. 2024. Can language models solve graph problems in natural language?. In Advances in Neural Information Processing Systems

  51. [59]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. In Proceedings of the 42nd international ACM SIGIR conference on Research and development in Information Retrieval

  52. [60]

    Zhen Wang, Jianwen Zhang, Jianlin Feng, and Zheng Chen. 2014. Knowledge graph embedding by translating on hyperplanes. In Proceedings of the AAAI conference on artificial intelligence

  53. [61]

    Zhihao Wen and Yuan Fang. 2024. Prompt tuning on graph-augmented low- resource text classification. IEEE Transactions on Knowledge and Data Engineering (2024)

  54. [62]

    Yaochen Xie, Zhao Xu, and Shuiwang Ji. 2022. Self-supervised representation learning via latent graph prediction. In International Conference on Machine Learning

  55. [63]

    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 Comprehensive Study on Text- attributed Graphs: Benchmarking and Rethinking. In Advance...

  56. [64]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al . 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115 (2024)

  57. [65]

    Zhilin Yang, William Cohen, and Ruslan Salakhudinov. 2016. Revisiting semi- supervised learning with graph embeddings. In Proceedings of the 33rd Interna- tional Conference on Machine Learning (ICML)

  58. [66]

    Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, Yongfeng Zhang, et al

  59. [67]

    Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. 2021. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing (2021)

  60. [68]

    Hengrui Zhang, Qitian Wu, Junchi Yan, David Wipf, and Philip S Yu. 2021. From canonical correlation analysis to self-supervised graph neural networks. In Ad- vances in Neural Information Processing Systems

  61. [69]

    Haiteng Zhao, Shengchao Liu, Chang Ma, Hannan Xu, Jie Fu, Zhihong Deng, Lingpeng Kong, and Qi Liu. 2023. GIMLET: A Unified Graph-Text Model for Instruction-Based Molecule Zero-Shot Learning. In Advances in Neural Informa- tion Processing Systems

  62. [70]

    Jianan Zhao, Le Zhuo, Yikang Shen, Meng Qu, Kai Liu, Michael Bronstein, Zhaocheng Zhu, and Jian Tang. 2023. Graphtext: Graph reasoning in text space. arXiv preprint arXiv:2310.01089 (2023)

  63. [71]

    Lei Zhu, Fangyun Wei, and Yanye Lu. 2024. Beyond text: Frozen large language models in visual signal comprehension. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . Table 8: Hyperparameter configurations for model pre- training on different...

  64. [2000]

    Information Retrieval (2000)

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

  65. [2015]

    In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

    Image-based recommendations on styles and substitutes. In Proceedings of the 38th international ACM SIGIR conference on research and development in information retrieval

  66. [2019]

    In The world wide web conference

    Graph neural networks for social recommendation. In The world wide web conference

  67. [2023]

    arXiv preprint arXiv:2308.07134 (2023)

    Natural language is all a graph needs. arXiv preprint arXiv:2308.07134 (2023)

  68. [2024]

    In International Conference on Machine Learning,

    LLaGA: Large Language and Graph Assistant. In International Conference on Machine Learning,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.