Pith. sign in

REVIEW 3 major objections 6 minor 62 references

CoEvoT claims that chain-of-thought reasoning works better on graphs when each intermediate thought rewrites the graph's token state, and reports consistent zero-shot gains over state-of-the-art graph–LLM methods on eight datasets.

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

2026-08-02 14:34 UTC pith:U7NNDAWE

load-bearing objection CoEvoT's closed-loop rewriting is a real idea and the ablation holds up, but the zero-shot headline numbers are undercut by tuning K, alpha, and s on the target test sets. the 3 major comments →

arxiv 2607.14114 v1 pith:U7NNDAWE submitted 2026-05-08 cs.CL cs.AI

CoEvoT: Co-Evolving Chain-of-Thought Prompting for Graph-LLM Reasoning

Haohua Niu , Xingtong Yu , Yang Liu , Junfeng Fang , Xuanting Xie , Jie Tan , Zhongjian Zhang , Hong Cheng
show 1 more author
Yuan Fang
This is my paper
classification cs.CL cs.AI
keywords co-evolving chain-of-thoughtgraph–LLM reasoningzero-shot cross-dataset transfergraph token rewritingcondition networknode classificationlink predictiontext-attributed graphs
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.

CoEvoT is a prompting framework for graph–large-language-model reasoning under distribution shift. Its central claim is that if intermediate chain-of-thought steps are allowed to rewrite the graph representation itself, rather than reasoning over a fixed graph snapshot, the model transfers better to unseen graphs with no labels. The framework alternates between two operations in a closed loop: the LLM produces a thought given the current graph-token state, and a small condition network converts that thought into node-specific residual updates to the graph embeddings; the updated tokens are then re-injected into the next prompt. On eight citation and e-commerce datasets, the paper reports consistent improvements over state-of-the-art graph–LLM baselines for zero-shot node classification and for link prediction transferred from a model trained only on node classification. The sympathetic reading: structural evidence should be a state that reasoning refines, not static context.

Core claim

On the paper's own terms, CoEvoT establishes that co-evolving CoT prompting — where each intermediate textual thought updates the graph-token evidence state via a lightweight condition network, and the updated tokens feed back into the next instruction — yields stronger zero-shot cross-dataset transfer than reasoning over a fixed graph. The evidence is a consistent lead over strong baselines on eight datasets spanning citation and e-commerce graphs, with the largest gains in the e-commerce setting where the distribution shift is larger. The ablation isolates the mechanism: multi-step reasoning with static graph tokens trails the full co-evolving model on essentially every dataset, so the gai

What carries the argument

The central object is the closed-loop co-evolving CoT cycle. At each step k, the frozen LLM is prompted with the current graph-token state and accumulated thoughts; the representation of the new thought t^k_v is passed through CondNet, a two-layer MLP hypernetwork, which outputs a node-specific conditional prompt p^k_v; the graph embedding is updated residually as h^{k+1}_v = h^k_v + α·p^k_v and re-projected into LLM token space. This mechanism is what carries the argument: it turns an intermediate thought from a textual trace into an actionable control signal that reshapes structural evidence, while the reshaped evidence in turn grounds the next reasoning step. Only the linear projector and

Load-bearing premise

A two-layer MLP trained only on source labels can map LLM-thought embeddings into node-specific residual updates that remain useful on unseen target graphs with different structure, semantics, and label spaces; the paper offers no analysis of when this learned mapping breaks, and the authors note in their limitations that gains may not transfer to graphs with weaker text signals.

What would settle it

At inference, replace the LLM's intermediate thoughts with random or fixed text (or random vectors in the LLM embedding space) and re-measure zero-shot accuracy. If accuracy does not drop relative to true thoughts, the thought-conditioned rewriting is not doing the state-refinement work claimed; if accuracy collapses, the claim is corroborated. A second check: shuffle the pairing between thoughts and nodes when training CondNet on the source graph — if cross-dataset accuracy stays high, the residual updates are not exploiting thought content.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Zero-shot node classification improves on all reported target datasets over the strongest graph–LLM baselines, with the largest gains where the structural and semantic shift is largest.
  • The learned rewriting loop transfers across tasks: a model adapted only for node classification also improves link prediction AUC on all datasets without task-specific training.
  • Multi-step reasoning alone is not enough; the ablation shows the text-to-graph rewriting loop is the component responsible for most of the gain.
  • Two reasoning steps and a moderate update weight (α ≈ 0.4) are sufficient, so the co-evolution loop remains cheap and predictable.
  • Inference overhead grows roughly linearly with the number of reasoning steps, making the cost of the closed loop controllable in practice.

Where Pith is reading between the lines

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

  • The same condition-network design could apply beyond graphs: any latent evidence state — retrieved passages, memory slots, or intermediate image features — could be rewritten by a model's own reasoning, suggesting a general 'reasoning rewrites evidence' pattern worth testing in other modalities.
  • A natural stress test the paper does not report: corrupting or randomizing the intermediate thoughts at inference. If performance survives, the rewriting loop is not actually reading thought content; if performance collapses, the paper's mechanism is confirmed.
  • Transfer success likely depends on the source label space being rich enough for CondNet to learn a generally useful thought-to-update map; on weakly text-attributed graphs or very different label spaces, the same learned updates may push target embeddings in source-specific directions — a boundary the authors themselves flag.

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 / 6 minor

Summary. The paper proposes CoEvoT, a graph–LLM prompting framework that couples multi-step chain-of-thought (CoT) reasoning with iterative rewriting of graph tokens. A frozen graph encoder produces node embeddings that are projected into LLM token space; at each of K reasoning steps, the LLM generates an intermediate thought, a two-layer condition network CondNet maps accumulated thought embeddings to node-specific residual updates h^{k+1}=h^k+αp^k, and the updated tokens are fed into the next instruction. Only the linear projector and CondNet are optimized on a labeled source graph; inference is then applied to target graphs. Experiments on citation and e-commerce datasets report node-classification accuracy, Macro-F1, link-prediction AUC, ablations against TEA-GLM and a no-rewriting variant, hyperparameter analyses, and inference-time comparisons.

Significance. The co-evolving idea is a simple and plausible extension of TEA-GLM and GCoT, and if validated it would be a useful contribution to label-efficient graph–LLM reasoning. The paper is transparent about using the released TEA-GLM encoder, explicitly identifies Variant 1 with TEA-GLM, and provides an algorithm and complexity analysis. The ablations consistently show the full mechanism outperforming the no-rewriting variant, which supports the internal mechanism. However, the headline zero-shot transfer claim is not currently established because target test sets are used for hyperparameter selection in §5.4, and the link-prediction table mixes source and target graphs. These issues are fixable within the scope of the manuscript, but they require additional experiments and a tightened evaluation protocol.

major comments (3)
  1. [§5.4 and Figs. 3–5; Tables 1–3] Section 5.4 sweeps K, α, and s using accuracy/AUC curves on Pubmed, History, Children, and Photo—the same target test datasets appearing in Tables 1–3—and then fixes defaults K=2, α=0.4, s=512 for all reported results. This leaks target labels into model selection and contradicts the stated protocol in §5.1 that the model is applied 'directly to unseen target datasets without further tuning.' The gains over TEA-GLM and GOFA may therefore be partly attributable to oracle hyperparameter selection rather than the co-evolution mechanism. Please re-run the main tables with hyperparameters selected on source-only validation (or fixed a priori) and report both settings; at minimum, show that the same defaults are optimal under source-only selection.
  2. [Table 2 and §5.1] Table 2 includes Arxiv and Computer as columns, but §5.1 designates Arxiv and Computer as the source datasets for training. The surrounding text claims evaluation on 'unseen target datasets' and 'cross-task link prediction'; including source graphs in the same table conflates cross-task transfer on the training domain with cross-dataset transfer. Please separate source and target results, clearly label them, and report variance or statistical tests for link prediction, which currently appears as a single run without error bars despite being part of the main claim.
  3. [§4.3, Eq. (5)–(6), Algorithm 1 lines 21–22; Appendix A] The method assumes that a two-layer CondNet MLP trained only on source labels maps LLM-thought embeddings to residual updates that remain useful on unseen target graphs with different structure, semantics, and label distributions. Appendix A concedes that gains may not transfer to graphs with weaker text signals, but the paper provides no analysis of when this transfer breaks. Since this assumption is the mechanism responsible for the reported cross-dataset gains, please add a stress test (e.g., text-perturbed graphs, graphs with non-textual features, or varying semantic alignment) or explicitly temper the cross-dataset claim to the evaluated settings.
minor comments (6)
  1. [Author affiliations] Typo: 'Telecommunieations' should be 'Telecommunications'.
  2. [References] References [19] and [20] are malformed: the venue/title fields are missing or incorrectly placed. Please fix the BibTeX entries.
  3. [Figs. 3–5] The y-axes in these hyperparameter figures use truncated and inconsistent ranges (e.g., 56–64 for some panels and 88–92 for others). Showing full scales and marking the chosen default values would make the saturation/instability claims easier to verify.
  4. [§5.1 GOFA reproduction] The text says GOFA results are the 'best-performing results after hyperparameter tuning.' Please clarify whether this tuning was performed on target test sets; if so, this should be disclosed and handled consistently with the paper's own hyperparameter protocol.
  5. [Checklist item 8 and Table 6] The paper's own checklist marks the compute-resources question as [No]. Table 6 reports inference times but does not specify the GPU model, memory, or total training compute; adding this information would improve reproducibility.
  6. [Table 5] On Children, COEVOT's Macro-F1 (0.255±0.008) overlaps with TEA-GLM's (0.252±0.005) within one standard deviation. The text's claim that COEVOT 'consistently achieves the best' Macro-F1 should be qualified or supported with significance tests.

Circularity Check

0 steps flagged

No definitional circularity; minor non-load-bearing self-citations only. Main caveat is target-set hyperparameter selection, a benchmarking concern rather than a circular reduction.

full rationale

CoEvoT is an empirical architecture paper rather than a derivation; the core equations (5)-(6) define a residual rewriting loop h^{k+1}=h^k+α·CondNet(t_1⊕...⊕t_k), which is not equivalent to its inputs by construction. Training uses only source labels (Eq. 7) with frozen encoder/LLM, and the reported gains are benchmarked against external baselines (TEA-GLM, GOFA, GCoT, etc.). Variant 1 is explicitly identified as TEA-GLM, so the comparison is transparent rather than a renamed result. Self-citations (GCoT [2], GraphPrompt-related works, [49]) are contextual or baseline references and are not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work. The reuse of TEA-GLM's released checkpoint is disclosed in §5.1 and Appendix F. The main validity concern is that §5.4 selects K=2, α=0.4, s=512 by sweeping target test datasets (Pubmed, History, Children, Photo), which conflicts with §5.1's 'without additional tuning' and weakens the zero-shot claim; however, this is a protocol/reproducibility issue rather than a definitional circularity, because no fitted parameter is relabeled as a prediction and no equation reduces to target labels. Appendix A's stated limitation on weak-text graphs is honest. Overall: no significant circularity; score reflects minor non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on a small number of fitting choices (α, K, s) plus a strong transfer assumption about CondNet. No new physical or formal entities are postulated. The main uncharged cost is that the hyperparameters were selected on the same target datasets used for the headline zero-shot numbers.

free parameters (4)
  • condition weight α = 0.4
    Set in §5.4/Fig. 4 by scanning values on target datasets Pubmed, History, Children, Photo for node classification and link prediction; the reported 'zero-shot' results then use this value.
  • number of reasoning steps K = 2
    Chosen in §5.4/Fig. 3 from target-dataset accuracy/AUC curves; K=2 is then used for all reported experiments.
  • CondNet hidden dimension s = 512
    Selected in §5.4/Fig. 5 on the same target datasets; adopted in all experiments.
  • training schedule hyperparameters = 60 epochs, batch 512, lr 2e-2; projector 2 epochs, batch 2, lr 1e-3
    Implementation choices in Appendix F; not swept on a validation set, but they affect the reported numbers.
axioms (4)
  • domain assumption Textual node attributes are informative enough on source and target graphs for the transfer task.
    The instruction templates in §4.2 and Appendix D inject node text (title/abstract) as the main semantic signal; Appendix A concedes gains may not transfer to graphs with weaker text signals.
  • domain assumption An LLM's intermediate thought representation can be mapped by a small MLP into a useful node-specific update in graph-embedding space.
    Eqs. (5)–(6) and Algorithm 1 line 21 assume CondNet transferability across datasets; no formal justification or failure analysis is given.
  • domain assumption TEA-GLM's pre-trained graph encoder checkpoint provides a valid shared initialization for both the baseline and COEVOT.
    §5.1 says COEVOT adapts the released TEA-GLM checkpoint; the paper does not actually run its own pre-training objective from §3.
  • domain assumption The contrastive alignment of graph embeddings with the LLM token-embedding PCA space is a sound semantic coordinate system.
    Appendix C builds the pre-training loss on PCA directions of LLM token embeddings; this is adopted from prior work and not independently validated here.

pith-pipeline@v1.3.0-alltime-deepseek · 24105 in / 9532 out tokens · 104820 ms · 2026-08-02T14:34:33.154287+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of CoEvoT: Co-Evolving Chain-of-Thought Prompting for Graph-LLM Reasoning." pith.science (2026). https://pith.science/paper/U7NNDAWE

@misc{pith2026260714114,
  author       = {Pith},
  title        = {Pith review of: CoEvoT: Co-Evolving Chain-of-Thought Prompting for Graph-LLM Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7NNDAWE}},
  note         = {Machine review of arXiv:2607.14114}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Graph learning under distribution shift presents a persistent challenge, where models adapt to new graphs with limited or even no supervision. Recent graph--LLM approaches move toward label-efficient prediction by linearizing graphs into prompts and using large language models (LLMs) as predictors, and can adopt Chain-of-Thought (CoT) prompting to exploit LLM's multi-step reasoning capability. However, existing CoT-based graph--LLM methods generate intermediate thoughts while conditioning on fixed graph tokens, limiting step-wise refinement of structural cues. In this paper, we propose CoEvoT, a simple yet effective co-evolving CoT prompting framework for graph--LLM reasoning. CoEvoT couples text-to-graph token rewriting and graph-to-text reasoning guidance in a closed loop: each intermediate textual thought is used to update the graph token evidence state via a lightweight condition network, and the updated tokens are fed back into the next-step instruction to guide subsequent LLM reasoning. This enables step-wise, state-aware evidence refinement, rather than reasoning over a fixed graph snapshot. Extensive experiments on eight datasets demonstrate that CoEvoT consistently outperforms state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2607.14114 by Haohua Niu, Hong Cheng, Jie Tan, Junfeng Fang, Xingtong Yu, Xuanting Xie, Yang Liu, Yuan Fang, Zhongjian Zhang.

Figure 1
Figure 1. Figure 1: Comparison of existing graph–LLM methods, GCOT and the proposed COEVOT. Although existing graph–LLM methods can in￾corporate CoT prompting [21, 17, 18], their rea￾soning remains largely language-only: interme￾diate thoughts are typically triggered by arti￾ficial instructions (e.g., “think step by step”), while the graph tokens used to construct the LLM instruction is fixed throughout reasoning. This static… view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of COEVOT. Graph tokenization. To maintain a stable interface to the LLM, a lightweight linear projector is used to map graph embeddings H to the LLM embedding space: Htoken = Linear(H; ϕ), (4) where ϕ is the learnable parameter. Htoken ∈ R |V |×L represents graph token with each row corre￾sponding to a node token of dimension L, matching the LLM token dimension. The node tokens serve as … view at source ↗
Figure 3
Figure 3. Figure 3: Effect of total inference steps K. 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 60 65 90 α Node classification Pubmed History Children Photo 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 54 55 62 63 64 α Link Prediction [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of hidden dimension s. Hidden dimension of condition network. We implement the condition network as a two-layer MLP and vary its hidden dimension s, with the results reported in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of step-wise node representation evolution on ArXiv. [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of step-wise node representation evolution on History. [PITH_FULL_IMAGE:figures/full_fig_p022_7.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

62 extracted references · 3 linked inside Pith

  1. [1]

    Local homophily-aware graph neural network with adaptive polynomial filters for scalable graph anomaly detection

    Zengyi Wo, Minglai Shao, Shiyu Zhang, and Ruijie Wang. Local homophily-aware graph neural network with adaptive polynomial filters for scalable graph anomaly detection. InSIGKDD, page 3180–3191, 2025

    2025

  2. [2]

    Gcot: Chain-of- thought prompt learning for graphs

    Xingtong Yu, Chang Zhou, Zhongwei Kuai, Xinming Zhang, and Yuan Fang. Gcot: Chain-of- thought prompt learning for graphs. InSIGKDD, pages 3669–3679, 2025

    2025

  3. [3]

    Gcc: Graph contrastive coding for graph neural network pre-training

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

    2020

  4. [4]

    D-tracker: Modeling interest diffusion in social activity tensor data streams

    Shingo Higashiguchi, Yasuko Matsubara, Koki Kawabata, Taichi Murayama, and Yasushi Sakurai. D-tracker: Modeling interest diffusion in social activity tensor data streams. In SIGKDD, page 460–471, 2025

    2025

  5. [5]

    E-commerce search via content collaborative graph neural network

    Guipeng Xv, Chen Lin, Wanxian Guan, Jinping Gou, Xubin Li, Hongbo Deng, Jian Xu, and Bo Zheng. E-commerce search via content collaborative graph neural network. InSIGKDD, page 2885–2897, 2023

    2023

  6. [6]

    Gppt: Graph pre-training and prompt tuning to generalize graph neural networks

    Mingchen Sun, Kaixiong Zhou, Xin He, Ying Wang, and Xin Wang. Gppt: Graph pre-training and prompt tuning to generalize graph neural networks. InSIGKDD, pages 1717–1727, 2022

    2022

  7. [7]

    Graph neural networks for multimodal single-cell data integration

    Hongzhi Wen, Jiayuan Ding, Wei Jin, Yiqi Wang, Yuying Xie, and Jiliang Tang. Graph neural networks for multimodal single-cell data integration. InSIGKDD, page 4153–4163, 2022

    2022

  8. [8]

    Groot: Effective design of biological sequences with limited experimental data

    Thanh VT Tran, Nhat Khang Ngo, Viet Anh Nguyen, and Truong Son Hy. Groot: Effective design of biological sequences with limited experimental data. InSIGKDD, pages 1385–1396, 2025

    2025

  9. [9]

    Semi-supervised classification with graph convolutional networks

    TN Kipf. Semi-supervised classification with graph convolutional networks. InICLR, 2017

    2017

  10. [10]

    Graph attention networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. InICLR, 2018

    2018

  11. [11]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. InNeurIPS, volume 30, 2017

    2017

  12. [12]

    Deep graph infomax

    Petar Velickovic, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. Deep graph infomax. InICLR, 2019

    2019

  13. [13]

    In Graph contrastive learning with augmentations, volume 33, pages 5812–5823, 2020

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

    2020

  14. [14]

    GraphPrompt: Unifying pre-training and downstream tasks for graph neural networks

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

    2023

  15. [15]

    Generalized graph prompt: Toward a unification of pre-training and downstream tasks on graphs.IEEE TKDE, 2024

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

    2024

  16. [16]

    Few-shot learning on graphs: from meta-learning to pre-training and prompting.arXiv preprint arXiv:2402.01440, 2024

    Xingtong Yu, Yuan Fang, Zemin Liu, Yuxia Wu, Zhihao Wen, Jianyuan Bo, Xinming Zhang, and Steven CH Hoi. Few-shot learning on graphs: from meta-learning to pre-training and prompting.arXiv preprint arXiv:2402.01440, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Llaga: Large language and graph assistant

    Runjin Chen, Tong Zhao, Ajay Jaiswal, Neil Shah, and Zhangyang Wang. Llaga: Large language and graph assistant. InICML, 2024

    2024

  18. [18]

    Graphgpt: Graph instruction tuning for large language models

    Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Lixin Su, Suqi Cheng, Dawei Yin, and Chao Huang. Graphgpt: Graph instruction tuning for large language models. InSIGIR, pages 491–500, 2024

    2024

  19. [19]

    InChain-of-thought prompting elicits reasoning in large language models, volume 35, pages 24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. InChain-of-thought prompting elicits reasoning in large language models, volume 35, pages 24824–24837, 2022. 10

    2022

  20. [20]

    InSelf-consistency improves chain of thought reasoning in language models, 2023

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

    2023

  21. [21]

    Graph chain-of-thought: Augmenting large language models by reasoning on graphs

    Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Suhang Wang, Yu Meng, and Jiawei Han. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. InACL, 2024

    2024

  22. [22]

    Extending the design space of graph neural networks by rethinking folklore weisfeiler-lehman

    Jiarui Feng, Lecheng Kong, Hao Liu, Dacheng Tao, Fuhai Li, Muhan Zhang, and Yixin Chen. Extending the design space of graph neural networks by rethinking folklore weisfeiler-lehman. InNeurIPS, volume 36, 2023

    2023

  23. [23]

    Mag- gnn: Reinforcement learning boosted graph neural network

    Lecheng Kong, Jiarui Feng, Hao Liu, Dacheng Tao, Yixin Chen, and Muhan Zhang. Mag- gnn: Reinforcement learning boosted graph neural network. InNeurIPS, volume 36, pages 12000–12021, 2023

    2023

  24. [24]

    GNNBoundary: Towards explaining graph neural networks through the lens of decision boundaries

    Xiaoqi Wang and Han Wei Shen. GNNBoundary: Towards explaining graph neural networks through the lens of decision boundaries. InICLR, 2024

    2024

  25. [25]

    Improving graph neural networks by learning continuous edge directions

    Seong Ho Pahng and Sahand Hormoz. Improving graph neural networks by learning continuous edge directions. InICLR, 2024

    2024

  26. [26]

    Graph learning with distributional edge layouts

    Xinjian Zhao, Chaolong Ying, Yaoyao Xu, and Tianshu Yu. Graph learning with distributional edge layouts. InSIGKDD, pages 2055–2066, 2025

    2055

  27. [27]

    Graph neural networks use graphs when they shouldn’t

    Maya Bechler-Speicher, Ido Amos, Ran Gilad-Bachrach, and Amir Globerson. Graph neural networks use graphs when they shouldn’t. InICML, 2024

    2024

  28. [28]

    How powerful are k-hop message passing graph neural networks

    Jiarui Feng, Yixin Chen, Fuhai Li, Anindya Sarkar, and Muhan Zhang. How powerful are k-hop message passing graph neural networks. InNeurIPS, volume 35, 2022

    2022

  29. [29]

    How powerful are graph neural networks? InICLR, 2018

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? InICLR, 2018

    2018

  30. [30]

    On the bottleneck of graph neural networks and its practical implications

    Uri Alon and Eran Yahav. On the bottleneck of graph neural networks and its practical implications. InICLR, 2021

    2021

  31. [31]

    Deep graph infomax

    Petar Veliˇckovi´c, William Fedus, William L Hamilton, Pietro Liò, Yoshua Bengio, and R Devon Hjelm. Deep graph infomax. InICLR, 2019

    2019

  32. [32]

    Graph contrastive learning automated

    Yuning You, Tianlong Chen, Yang Shen, and Zhangyang Wang. Graph contrastive learning automated. InICML, pages 12121–12132, 2021

    2021

  33. [33]

    Graphmae: Self-supervised masked graph autoencoders

    Zhenyu Hou, Xiao Liu, Yukuo Cen, Yuxiao Dong, Hongxia Yang, Chunjie Wang, and Jie Tang. Graphmae: Self-supervised masked graph autoencoders. InSIGKDD, pages 594–604, 2022

    2022

  34. [34]

    Unigraph: Learning a unified cross-domain foundation model for text-attributed graphs

    Yufei He, Yuan Sui, Xiaoxin He, and Bryan Hooi. Unigraph: Learning a unified cross-domain foundation model for text-attributed graphs. InSIGKDD, pages 448–459, 2025

    2025

  35. [35]

    Graph neural prompting with large language models

    Yijun Tian, Huan Song, Zichen Wang, Haozhu Wang, Ziqing Hu, Fang Wang, Nitesh V Chawla, and Panpan Xu. Graph neural prompting with large language models. InAAAI, volume 38, pages 19080–19088, 2024

    2024

  36. [36]

    Large language model meets graph neural network in knowledge distillation

    Shengxiang Hu, Guobing Zou, Song Yang, Shiyi Lin, Yanglan Gan, Bofeng Zhang, and Yixin Chen. Large language model meets graph neural network in knowledge distillation. InAAAI, volume 39, pages 17295–17304, 2025

    2025

  37. [37]

    Stage: Simplified text-attributed graph embeddings using pre-trained llms

    Aaron Zolnai-Lucas, Jack Boylan, Chris Hokamp, and Parsa Ghaffari. Stage: Simplified text-attributed graph embeddings using pre-trained llms. InKaLLM, 2024

    2024

  38. [38]

    Login: A large language model consulted graph neural network training framework

    Yiran Qiao, Xiang Ao, Yang Liu, Jiarong Xu, Xiaoqian Sun, and Qing He. Login: A large language model consulted graph neural network training framework. InWSDM, pages 232–241, 2025. 11

    2025

  39. [39]

    Graph prompt learning: A comprehensive survey and beyond.arXiv preprint arXiv:2311.16534, 2023

    Xiangguo Sun, Jiawen Zhang, Xixi Wu, Hong Cheng, Yun Xiong, and Jia Li. Graph prompt learning: A comprehensive survey and beyond.arXiv preprint arXiv:2311.16534, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Universal prompt tuning for graph neural networks

    Taoran Fang, Yunchao Zhang, Yang Yang, Chunping Wang, and Lei Chen. Universal prompt tuning for graph neural networks. InNeurIPS, volume 36, 2023

    2023

  41. [41]

    Graphtext: Graph reasoning in text space

    Jianan Zhao, Le Zhuo, Yikang Shen, Meng Qu, Kai Liu, Michael Bronstein, Zhaocheng Zhu, and Jian Tang. Graphtext: Graph reasoning in text space. InNeurIPS, 2024

    2024

  42. [42]

    Graphtranslator: Aligning graph model to large language model for open-ended tasks

    Mengmei Zhang, Mingwei Sun, Peng Wang, Shen Fan, Yanhu Mo, Xiaoxiao Xu, Hong Liu, Cheng Yang, and Chuan Shi. Graphtranslator: Aligning graph model to large language model for open-ended tasks. InWWW, pages 1003–1014, 2024

    2024

  43. [43]

    Graph2text or graph2token: A perspective of large language models for graph learning.ACM, 2026

    Shuo Yu, Yingbo Wang, Ruolin Li, Guchun Liu, Yanming Shen, Shaoxiong Ji, Bowen Li, Fengling Han, Xiuzhen Zhang, and Feng Xia. Graph2text or graph2token: A perspective of large language models for graph learning.ACM, 2026

    2026

  44. [44]

    The graph neural network model.IEEE transactions on neural networks, 20(1):61–80, 2008

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model.IEEE transactions on neural networks, 20(1):61–80, 2008

    2008

  45. [45]

    Augmenting low-resource text classification with graph-grounded pre-training and prompting

    Zhihao Wen and Yuan Fang. Augmenting low-resource text classification with graph-grounded pre-training and prompting. InSIGIR, pages 506–516, 2023

    2023

  46. [46]

    Graphclip: Enhancing transferability in graph foundation models for text-attributed graphs

    Yun Zhu, Haizhou Shi, Xiaotang Wang, Yongchao Liu, Yaoke Wang, Boci Peng, Chuntao Hong, and Siliang Tang. Graphclip: Enhancing transferability in graph foundation models for text-attributed graphs. InWWW, pages 2183–2197, 2025

    2025

  47. [47]

    Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings

    Duo Wang, Yuan Zuo, Fengzhi Li, and Junjie Wu. Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings. InNeurIPS, volume 37, pages 5950–5973, 2024

    2024

  48. [48]

    Hypernetworks.arXiv preprint arXiv:1609.09106, 2016

    David Ha, Andrew Dai, and Quoc V Le. Hypernetworks.arXiv preprint arXiv:1609.09106, 2016

    Pith/arXiv arXiv 2016

  49. [49]

    Node-time conditional prompt learning in dynamic graphs

    Xingtong Yu, Zhenghao Liu, Xinming Zhang, and Yuan Fang. Node-time conditional prompt learning in dynamic graphs. InICLR, 2025

    2025

  50. [50]

    Open graph benchmark: Datasets for machine learning on graphs

    Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs. InNeurIPS, volume 33, pages 22118–22133, 2020

    2020

  51. [51]

    Harnessing explanations: Llm-to-lm interpreter for enhanced text-attributed graph representation learning

    Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, and Bryan Hooi. Harnessing explanations: Llm-to-lm interpreter for enhanced text-attributed graph representation learning. InICLR, 2024

    2024

  52. [52]

    A comprehensive study on text-attributed graphs: Benchmarking and rethinking

    Hao Yan, Chaozhuo Li, Ruosong Long, Chao Yan, Jianan Zhao, Wenwen Zhuang, Jun Yin, Peiyan Zhang, Weihao Han, Hao Sun, et al. A comprehensive study on text-attributed graphs: Benchmarking and rethinking. InNeurIPS, volume 36, pages 17238–17264, 2023

    2023

  53. [53]

    Multilayer perceptron (mlp)

    Hind Taud and Jean-Franccois Mas. Multilayer perceptron (mlp). InGeomatic approaches for modeling land change scenarios, pages 451–455. 2017

    2017

  54. [54]

    Geometric knowledge distillation: Topology compression for graph neural networks

    Chenxiao Yang, Qitian Wu, and Junchi Yan. Geometric knowledge distillation: Topology compression for graph neural networks. InNeurIPS, volume 35, pages 29761–29775, 2022

    2022

  55. [55]

    Global-local graph neural networks for node-classification

    Moshe Eliasof and Eran Treister. Global-local graph neural networks for node-classification. Pattern Recognition Letters, 184:103–110, 2024

    2024

  56. [56]

    Nodeformer: A scalable graph structure learning transformer for node classification

    Qitian Wu, Wentao Zhao, Zenan Li, David P Wipf, and Junchi Yan. Nodeformer: A scalable graph structure learning transformer for node classification. InNeurIPS, volume 35, pages 27387–27401, 2022

    2022

  57. [57]

    Difformer: Scalable (graph) transformers induced by energy constrained diffusion

    Qitian Wu, Chenxiao Yang, Wentao Zhao, Yixuan He, David Wipf, and Junchi Yan. Difformer: Scalable (graph) transformers induced by energy constrained diffusion. InICLR, 2023. 12

    2023

  58. [58]

    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. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, 2023

    2023

  59. [59]

    One for all: Towards training one graph model for all classification tasks

    Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang. One for all: Towards training one graph model for all classification tasks. InICLR, 2024

    2024

  60. [60]

    Gofa: A generative one-for-all model for joint graph language modeling

    Lecheng Kong, Jiarui Feng, Hao Liu, Chengsong Huang, Jiaxin Huang, Yixin Chen, and Muhan Zhang. Gofa: A generative one-for-all model for joint graph language modeling. InICLR, 2025. 13 A Limitations Although COEVOT improves transfer performance, it introduces additional inference overhead because graph evidence is updated across multiple reasoning steps. ...

    2025

  61. [62]

    Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or ...

  62. [256]

    Children

    The number of inference steps is set to 2. The condition-net is implemented as a two-layer MLP with a bottleneck architecture, where the input dimension is 256 and the hidden dimension is set to 32. GOFA setting.For GOFA, we set the optimizer to AdamW, with a learning rate of 1×10 −4, weight decay of 0.1, batch size of 8, dropout of 0.0, gradient clipping...