Agentic Graph Token Reasoning

Yi Yang; Zhuoyi Peng

REVIEW 2 major objections 5 minor 48 references

Agentic Graph Token Reasoning

T0 review · 2 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Letting a language model choose which graph views to encode at each reasoning step beats fixed single-shot graph tokens, with the graph-token medium itself driving the gain.

desk verdict Genuinely new agentic graph-token framework with a broad evaluation, but the headline medium-attribution claim is undercut by a possibly weak text baseline; worth refereeing, but treat the 29-point gap with caution. read the letter →

arxiv 2608.00542 v1 pith:KGNX2Z4E submitted 2026-08-01 cs.LG

Zhuoyi Peng , Yi Yang This is my paper

classification cs.LG

keywords agenticreasoninggraphtokensneuralnetworkslargelanguagemodelstext-attributedgraphszero-shottransferpreferenceoptimizationnodeclassification

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

Existing graph-token methods compress one predefined graph view into a block of soft tokens that the language model reads once, before it has even seen the question. This paper claims that this static design is the bottleneck: graph analysis should be agentic, with the model repeatedly choosing which part of the graph to encode next, reading the newly materialised graph tokens, and stopping when it has enough evidence. The paper builds Agentic Graph Token (AGT) reasoning and trains it in three stages—self-supervised token reading, trajectory fine-tuning with a consistency regulariser, and preference optimisation that rewards agreement between graph tokens and node text. Across seven graph domains AGT outperforms comparably sized graph-token and text-agent baselines, and a direct medium ablation shows why: the same agentic loop with verbalised text retains only 28.1% zero-shot accuracy on unseen domains against 57.4% with graph tokens. The conclusion a sympathetic reader takes away is that the graph-token medium, not the agentic loop alone, carries the gain.

What carries the argument

The central object is the agentic graph-token block: each step the policy emits an action from a small open set (the anchor node, 1- to 3-hop neighbourhoods, a retrieved set of similar nodes, a cluster, and an answer action), and a graph encoder compresses the named graph view into a fixed-length (128-token) continuous block that is spliced into the running context and read by the language model. Because the block length is constant, the context cost of evidence gathering is bounded. The training machinery is a three-stage pipeline: self-supervised objectives (text reconstruction and masked link prediction) that ground the tokens; trajectory SFT with a KL regulariser that forces the reading

What would settle it

Run the Stage 3 preference optimization with a relaxed filter that also keeps pairs where the inconsistent rollout is correct but differs (removing the correctness condition), and compare test accuracy; if the gain vanishes, the correctness-conditioned selection was the driver. Alternatively, at inference time replace the graph-token block with a same-shaped random token block in the zero-shot setting; if accuracy stays near 57%, the medium's content is not what transfers.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a reasoning step on a graph is best carried by a graph token, not by language. The agentic loop matters only insofar as it lets the model request the right graph views. The decisive evidence is the medium ablation of Section 5.5: the authors take their full agentic framework and swap only what each action returns—serialised text in one variant, a fixed-size graph-token block in the other—while keeping the action space, rollout, and training recipe identical. In-domain the two are close (81.0 vs 77.8 average), but on zero-shot transfer the graph-token medium retains 57.4% accuracy against 28.1% for the verbalised one, a 29-point gap. Error analysis shows

Load-bearing premise

The final training stage keeps a preference pair only when the consistent rollout is correct and the text-corrupted one is wrong, so the policy is shaped by the examples where graph grounding happened to succeed; if those cases are unrepresentative of hard test nodes, the reported gains may overstate the grounding mechanism.

Editorial extensions

If this is right

Graph-token methods should be redesigned as interactive, trajectory-dependent systems rather than single-pass encoders; the fixed-size token block keeps the added cost of each reasoning step constant.
Zero-shot transfer to new graph domains appears to depend on the compact structural code in graph tokens: verbalising the same neighbourhoods loses 29 points of accuracy on unseen targets.
The policy's action count is adaptive, so a caller can trade a small accuracy change for a shorter context by capping the budget near one to three actions without leaving the Pareto frontier.
Adding new graph-view actions (e.g., temporal or hypergraph scopes) is a drop-in extension, since the action space is open by design.

Reading between the lines

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

A testable extension of the medium claim: if graph tokens act as a transferable structural code, training the encoder on a diverse set of graph families should further widen the zero-shot gap over any text verbaliser.
The Stage 3 filter keeps only pairs where the consistent rollout is correct and the inconsistent one wrong; measuring the kept-pair rate by node difficulty would show whether the preference signal is concentrated on easy anchors, a question the paper does not report.
The same graph-text inconsistency principle could be applied to other token modalities (vision, audio) where models exhibit a 'text shortcut' over a rich perceptual input.

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.

Desk Editor's Note

read the letter

Dear [Colleague],

The paper to know about: it's the first agentic graph-token reasoning framework, and the empirical sweep is broad, but the headline claim—that the graph-token medium, not the agentic loop, drives the 29-point zero-shot gap—rests on a text baseline that may be underpowered. I'd send it to a referee, but the medium attribution needs proof that Ours-Text actually reads the serialized views.

What's genuinely new: prior graph-token methods are single-shot; prior agentic graph methods work over text. AGT unites them. The three-stage recipe—reading, consistency regularizer, preference optimization—is thoughtfully designed, and the medium-vs-text ablation in Table 5 is exactly the right control. The evaluation is unusually broad: seven in-domain datasets, seven zero-shot targets, ablations, and an error-structure analysis. That's real work.

The soft spots are real too. The stress-test note has teeth: the paper asserts serialized views fit in the 2048-token context but reports no token counts. A 3-hop neighborhood serialized as text can blow past that easily. More importantly, the trajectory corpus was synthesized from the graph-token policy's actions; the text model may be learning actions that don't map well to text, and there's no ablation showing Ours-Text's accuracy depends on the graph-view text. If the text model is ignoring those views, the loop is inactive in that arm, and the 29-point gap measures a broken baseline. The fact that Ours-Text zero-shot (28.1%) is below the strongest text agent baseline (AgentGL, 34.5%) is a red flag. It's not damning—in-domain the two media are close (77.8 vs 81.0), so the text model is doing something—but the medium attribution is not clean.

Second, Stage 3's preference-pair filter conditions on the ground-truth label: pairs are kept only when the consistent rollout is correct and the inconsistent one is wrong. That selects on outcomes and could bias the preference distribution; the paper doesn't report the keep rate. That's not circular with external benchmarks, but it's a real distribution-shift issue.

Third, no code, no checkpoints, no error bars or significance tests, and the zero-shot protocol for different label spaces is underspecified. For a paper making 29-point claims, that's too thin.

Still, these are fixable. The framework is novel, the writing is clear, and the core idea is plausible. I'd send it to peer review with a request for the missing artifacts and a demonstration that Ours-Text genuinely consumes the serialized graph views. If that holds, this is an important advance; if not, it's a solid framework with an overstated attribution.

Bring it to the reading group; it's a good case for discussing evaluation design.

Best,

[Your name]

Referee Report

2 major / 5 minor

Summary. The paper introduces Agentic Graph Token (AGT) reasoning, in which a language model interleaves generation with on-demand graph-token materializations: at each step the model emits an action selecting a graph view (node, k-hop neighborhood, retrieval set, cluster), a GNN encoder compresses that view into a fixed-length block of continuous tokens spliced into the context, and the loop terminates with an answer action. Training proceeds in three stages: self-supervised alignment for reading heterogeneous graph tokens, trajectory SFT with a consistency regularizer over perturbed graph tokens, and IPO preference optimization over rollouts that are consistent versus inconsistent between node text and graph tokens. Claims include: (i) large gains over same-backbone graph-token and agentic baselines on seven in-domain graph domains, (ii) zero-shot transfer to seven unseen domains, and (iii) the central attribution that the graph-token medium, not the agentic loop alone, drives the gains, supported by a text-medium ablation (Ours-Text vs. Ours-Token) in Table 5. The paper also includes analyses of error structure, inference cost, training cost, and sensitivity to hyperparameters.

Significance. If the central claim is established, AGT is a valuable step beyond static graph-token methods and text-based agentic approaches: it makes the graph-token representation part of the reasoning loop in a way that is genuinely new, and the three-stage recipe is concrete and reproducible in structure. The empirical evaluation is broad, with seven in-domain datasets, seven zero-shot targets, same-backbone re-implemented baselines, and detailed appendices on corpora, cost, and sensitivity. The paper is also honest about limitations (frozen text features, IPO pair types). The main risk is that the medium-attribution result, which is the paper's headline, rests on a text-medium ablation whose fairness and functional integrity are not demonstrated. A second risk is the ground-truth-conditioned preference filtering in Stage 3, which could make the reported gains partly an artifact of selection on training outcomes. With additional experiments, these concerns are addressable within the manuscript's scope.

major comments (2)

[§5.5, Table 5] The central claim that the graph-token medium, not the agentic loop, drives the gains rests entirely on Ours-Text being a functional, fairly configured text-based agent. The paper asserts 'the serialised views fit within that budget' but reports no token counts. On large graphs, a 3-hop neighborhood or a cluster serialized as text can easily exceed the 2048-token context unless aggressively truncated; truncation would penalize the text medium independently of representation quality. Moreover, the Stage 2 trajectory corpus (Appendix D) is synthesized from actions chosen under the graph-token policy (e.g., calling three_hop_token because a compact block resolves ambiguity); verbalizing those same actions may not benefit the text model in the same way. No analysis shows that Ours-Text actually reads and exploits the serialized graph views. If the text model ignores these views, the agentic
[§4.3, Eq. (10)] The Stage 3 preference-pair filter conditions on the ground-truth label: a pair is kept only when the consistent rollout is correct (Ac=y), the inconsistent rollout is wrong (Ai≠y), and the answers differ. This is selection on training outcomes. If the Stage 2 policy's errors are concentrated on hard or ambiguous nodes, the resulting preference distribution may not represent the test distribution, and the reported gains from Stage 3 (Table 4 shows non-negligible contributions, e.g., +5.1 AUC on PubMed LP and +1.9 NC average) may partly reflect selection rather than a robust grounding mechanism. The paper reports 'on the order of 300 pairs per run' but does not analyze the rate at which pairs are kept, nor the characteristics of kept versus discarded anchors. Please report the number of rollouts and kept pairs per dataset, and compare against a label-agnostic or unfiltered preference base

minor comments (5)

[General] The phrase 'graph token from one-shot preprocessing' in the conclusion has a grammar slip; 'turns graph token from' should likely be 'turns graph tokens from' or 'turns the graph token from'.
[Table 2] Some entries are missing spaces or are visually merged, e.g., '62.975.950.6' in the LLaGA row for CiteSeer NC. Formatting such numbers would improve readability.
[§5.2] The paper reports 3-run averages but no standard deviations or significance tests. Table 4 explicitly says several action-ablation differences are 'within measurement noise'; the same caution should apply to the small differences in Tables 2 and 5.
[§5.7] The claim that the policy 'issues on average 4 actions per trajectory' on ogbn-arxiv is not accompanied by a distribution or variance. Since trajectory length is a central adaptive property, a histogram or range would be informative.
[Appendix D] The Stage 3 pair count of 'on the order of 300 pairs per run' appears without a dataset-by-dataset breakdown. Given the selection concern in the major comment, this breakdown is important.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are supported by external baselines and controlled ablations rather than by construction or self-citation.

full rationale

The paper's main derivation chain is empirical and self-contained against external benchmarks. The central claim that the graph-token medium, rather than the agentic loop, drives gains is tested by a direct medium ablation (Table 5) that keeps the action space, rollout, and training recipe identical and changes only the representation returned by each action. The Stage 3 preference optimization filters rollouts using training labels (consistent answer correct, inconsistent answer wrong); this is a standard training-signal construction, not a fitted parameter renamed as a prediction, and it does not make the reported test accuracy true by construction. The authors' own limitations appendix (Appendix A) acknowledges design choices and future work, but none of these admissions amount to a circular step. There are no load-bearing self-citations: the cited baselines and related work are by other authors, and no uniqueness theorem or ansatz is imported from the authors' prior work to force the framework. The potential concern that the Ours-Text baseline in Table 5 may be underpowered (unverified context fit, no evidence it attends to serialized views) is an empirical validity question, not a circularity of the derivation chain. Overall, the predictions are not equivalent to the inputs by definition, and no quoted equation or fitted quantity reduces the claimed results to their own assumptions.

Assumptions & free parameters 7 free parameters · 5 assumptions · 2 invented entities

The central claim rests on a stack of design choices: a specific encoder architecture, a fixed action space, a frozen text embedder, and a three-stage training recipe with several hyperparameters (lambda_KL, p_ed, p_tm, omega, beta, block size). None of these are derived from first principles; they are justified empirically. The paper is honest about some of these (e.g., Appendix A notes the frozen mpnet limitation), but the lack of code and the absence of sensitivity analyses for the key hyperparameters mean a reader cannot yet tell which choices are load-bearing.

free parameters (7)

lambda_KL (Stage 2 consistency weight) = 0.5
Chosen by hand for the KL regularizer in Eq. 9; no sensitivity analysis reported in the main text, only a general statement in Appendix H about pooler choices, not this weight.
edge drop probability p_ed = 0.1
Chosen for the token-robust trajectory augmentation; no ablation shown for this value.
token mask probability p_tm = 0.2
Chosen for the token-mask augmentation in Stage 2; no ablation shown.
answer token weight omega = 20
Hand-picked for the token-weighted cross-entropy in Eq. 8; no sensitivity analysis.
IPO beta = 0.3
Chosen for the IPO objective in Eq. 10; no sensitivity analysis reported.
per-action token block size = 128
The default per-action graph-token block length. Appendix H shows accuracy saturates around 64-128 tokens, so this is a reasonable but hand-chosen default.
number of training nodes / pairs = 3000 train nodes, 1000 test nodes; 1500+1500 LP pairs
Follows AgentGL's protocol, but the choice of evaluation sample size (1000 test nodes) affects the statistical precision of all reported numbers.

assumptions (5)

domain assumption MPNet sentence embeddings provide a sufficient text-feature representation for all domains.
The paper uses frozen MPNet features throughout and notes in Appendix A that a stronger feature extractor might improve results. The success of the method depends on these features carrying enough semantic signal.
domain assumption The fixed action space (Table 1) is sufficient to express the graph evidence needed for the evaluated tasks.
The model can only request views in the action space. The paper acknowledges the space is not exhaustive, but the reported results are conditional on this specific set of actions.
domain assumption GraphSAGE pooling over the chosen graph view produces a block that preserves the discriminative signal needed for the task.
This is the core representational assumption of all graph-token methods; the paper provides empirical evidence but no formal guarantee. Appendix H shows some robustness to GNN backbone choice.
ad hoc to paper The trajectory SFT corpus is representative of the test-time distribution of questions and graphs.
Stage 2 trains on synthesized trajectories (Section 4.2). The synthesis rules are described at a high level in Appendix D, and their coverage of the space of possible reasoning paths is not analyzed.
standard math Identity Preference Optimisation (IPO) is a valid and stable objective for this small preference set.
IPO is a known method (cited as [1]); the paper uses it to avoid DPO's unbounded log-ratio on small pair sets. This is a reasonable but unproven modeling choice for this setting.

invented entities (2)

Agentic graph-token (AGT) block
purpose: A fixed-size block of continuous tokens produced on demand by a GNN encoder for a model-selected graph view, spliced into the LLM context to enable step-by-step graph reasoning.
This is the paper's central new abstraction. It is an architectural/representational device, not a physical entity. Its utility is evidenced only by the paper's own experiments; no external benchmark or independent implementation is provided.
Graph-token consistency regularizer (CRJT)
purpose: A training objective that aligns the LLM's response distributions over clean and perturbed graph-token blocks, forcing reliance on token content.
A training technique, not a standalone entity. Its effectiveness is shown only through the paper's own ablations.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Agentic Graph Token Reasoning." pith.science (2026). https://pith.science/paper/KGNX2Z4E

@misc{pith2026260800542,
  author       = {Pith},
  title        = {Pith review of: Agentic Graph Token Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KGNX2Z4E}},
  note         = {Machine review of arXiv:2608.00542}
}

read the original abstract

Graphs model relational data throughout science and industry, from citation networks to product co-purchase graphs. Because the nodes of many such graphs carry rich text, a growing line of work applies large language models (LLMs) to graph analysis. The most graph-native of these methods use graph tokens: a graph encoder compresses a graph view, such as a node, its k-hop neighbourhood, or a cluster, into a short block of continuous tokens that jointly encodes node attributes and topology and is read directly by the model. Existing methods, however, use graph tokens in a static single-shot manner: they encode one predefined graph view before the model has even seen the target and never revise it, leaving the model's step-by-step reasoning ability unused. We introduce agentic graph token reasoning, which recasts graph tokenization as part of the reasoning process itself. At each step, the model chooses which graph view to encode and at what granularity; a graph encoder is invoked on demand to materialise the corresponding graph tokens; and the resulting block is spliced into the running context. The model thus reasons step by step in the graph token space, and the tokens it reads are trajectory-dependent. We realise this with a three-stage training pipeline: (i) self-supervised tasks that teach the model to read heterogeneous graph tokens, (ii) a token-robust trajectory stage with a graph-token consistency regulariser, and (iii) preference optimisation that rewards trajectories in which the graph-token evidence and the node-text evidence agree. Across evaluations spanning seven graph domains, our models outperform a broad set of baselines by a large margin and transfer zero-shot to unseen domains without any per-target fine-tuning. More broadly, this work pushes LLM-based graph analysis from static graph-token encoders towards a graph-native agent paradigm.

Figures

Figures reproduced from arXiv: 2608.00542 by the authors.

**Figure 2.** Overview of Agentic Graph Token (AGT) reasoning. Stage 1 grounds the block with two self-supervised reading tasks; [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Error-fraction density over node degree: [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (8 more)

**Figure 4.** Figure 4: Accuracy versus inference tokens on ogbnproducts. Blue 3B, orange 7B; Star: ours, circle: baseline, and the dashed line is the Pareto frontier for each size. More results are in Appendix F. 5.8 Further analysis More studies are deferred to the appendix for space. Appe…

**Figure 5.** Figure 5: Error-fraction density over node degree, one panel per dataset, using the same conventions as Figure 3. Our density [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]

**Figure 6.** Figure 6: Accuracy against inference context (part 1 of 2): ogbn-arxiv, ogbn-products, PubMed, Reddit. Colour encodes backbone [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Accuracy against inference context (part 2 of 2): arXiv-2023, CiteSeer, STRING-db. Same conventions as Figure 6: blue [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

**Figure 9.** Figure 9: Per-stage training-time breakdown of our three [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

**Figure 10.** Figure 10: Per-action token budget vs. full-pipeline NC ac [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

**Figure 11.** Figure 11: GNN backbone vs. full-pipeline NC accuracy ( [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

**Figure 12.** Figure 12: Agentic trajectory for one ogbn-arxiv node and what a Stage-1 reader recovers from each injected graph token. The policy escalates from the paper’s own text to pooled community and neighbourhood tokens; decoding those tokens back to text shows they carry computer-visi…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 23 canonical work pages

[1]

Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoret- ical paradigm to understand learning from human preferences. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 4447–4455

2024
[2]

Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. 2024. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 17682–17690

2024
[3]

Runjin Chen, Tong Zhao, Ajay Jaiswal, Neil Shah, and Zhangyang Wang. 2024. LLaGA: large language and graph assistant. InProceedings of the 41st International Conference on Machine Learning. 7809–7823

work page 2024
[4]

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 Newsletter25, 2 (2024), 42–61

work page 2024
[5]

Zhikai Chen, Haitao Mao, Hongzhi Wen, Haoyu Han, Wei Jin, Haiyang Zhang, Hui Liu, and Jiliang Tang. 2024. Label-free node classification on graphs with large language models (llms). InInternational Conference on Learning Representa- tions, Vol. 2024. 31632–31655

work page 2024
[6]

Yuntian Deng, Yejin Choi, and Stuart Shieber. 2024. From explicit cot to implicit cot: Learning to internalize cot step by step.arXiv preprint arXiv:2405.14838 (2024)

arXiv 2024
[7]

Yi Fang, Dongzhe Fan, Daochen Zha, and Qiaoyu Tan. 2024. Gaugllm: Improving graph contrastive learning for text-attributed graphs with large language models. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 747–758

work page 2024
[8]

Sachin Goyal, Ziwei Ji, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar, and Vaishnavh Nagarajan. [n. d.]. Think before you speak: Training Language Models With Pause Tokens. InThe Twelfth International Conference on Learning Representations

work page

Show all 48 references

[9]

Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. 2025. DeepSeek-R1 incen- tivizes reasoning in LLMs through reinforcement learning.Nature645, 8081 (2025), 633–638

2025
[10]

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

2023 arXiv
[11]

Hamilton, Rex Ying, and Jure Leskovec

William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive Representation Learning on Large Graphs. InAdvances in Neural Information Processing Systems (NeurIPS)

2017
[12]

Shibo Hao, Sainbayar Sukhbaatar, DiJia Su, Xian Li, Zhiting Hu, Jason Weston, and Yuandong Tian. 2024. Training large language models to reason in a contin- uous latent space.arXiv preprint arXiv:2412.06769(2024)

2024 arXiv
[13]

Xiaoxin He, Xavier Bresson, Thomas Laurent, Adam Perold, Yann LeCun, and Bryan Hooi. 2023. Harnessing explanations: Llm-to-lm interpreter for enhanced text-attributed graph representation learning.arXiv preprint arXiv:2305.19523 (2023)

2023 arXiv
[14]

Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. 2020. Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems 33 (2020), 22118–22133

2020
[15]

Jin Huang, Xingjian Zhang, Qiaozhu Mei, and Jiaqi Ma. 2023. Can LLMs Effec- tively Leverage Graph Structural Information through Prompts, and Why?arXiv preprint arXiv:2309.16595(2023)

2023 arXiv
[16]

Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. Openai o1 system card.arXiv preprint arXiv:2412.16720(2024)

2024 arXiv
[17]

Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al . 2023. Mistral 7B.arXiv preprint arXiv:2310.06825(2023)

2023 arXiv
[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. InFindings of the Association for Computational Linguistics...

2024
[19]

Thomas N Kipf and Max Welling. 2016. Semi-supervised classification with graph convolutional networks.arXiv preprint arXiv:1609.02907(2016)

2016 arXiv
[20]

Lecheng Kong, Jiarui Feng, Hao Liu, Chengsong Huang, Jiaxin Huang, Yixin Chen, and Muhan Zhang. 2025. GOFA: A generative one-for-all model for joint graph language modeling. In13th International Conference on Learning Representations, ICLR 2025. International Conference on Lea...

2025
[21]

Rui Li, Jiwei Li, Jiawei Han, and Guoyin Wang. 2024. Similarity-based neighbor selection for graph llms.arXiv preprint arXiv:2402.03720(2024)

2024 arXiv
[22]

Jiajin Liu, Yuanfu Sun, Dongzhe Fan, and Qiaoyu Tan. 2026. GraphSearch: Agentic Search-Augmented Reasoning for Zero-Shot Graph Learning.arXiv preprint arXiv:2601.08621(2026)

2026
[23]

Zheyuan Liu, Xiaoxin He, Yijun Tian, and Nitesh V Chawla. 2024. Can we soft prompt llms for graph learning tasks?. InCompanion proceedings of the ACM web conference 2024. 481–484

2024
[24]

Rafael Rafailov, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D Manning, and Chelsea Finn. 2023. Direct preference optimization: your lan- guage model is secretly a reward model. InProceedings of the 37th International Conference on Neural Information Processing Sy...

2023
[25]

Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. InProceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCN...

2019
[26]

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300(2024)

2024 arXiv
[27]

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

2018 arXiv
[28]

Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, and Tie-Yan Liu. 2020. Mpnet: Masked and permuted pre-training for language understanding.Advances in neural information processing systems33 (2020), 16857–16867

2020
[29]

Yuanfu Sun, Kang Li, Dongzhe Fan, Jiajin Liu, and Qiaoyu Tan. 2026. Agentgl: Towards agentic graph learning with llms via reinforcement learning. InProceed- ings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 25313–25335

2026
[30]

Yuanfu Sun, Zhengnan Ma, Yi Fang, Jing Ma, and Qiaoyu Tan. 2025. Graphicl: Unlocking graph learning potential in llms through structured prompt design. InFindings of the Association for Computational Linguistics: NAACL 2025. 2440– 2459

2025
[31]

Damian Szklarczyk, Annika L Gable, David Lyon, Alexander Junge, Stefan Wyder, Jaime Huerta-Cepas, Milan Simonovic, Nadezhda T Doncheva, John H Morris, Peer Bork, et al. 2019. STRING v11: protein–protein association networks with in- creased coverage, supporting functional disc...

2019
[32]

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

2024
[33]

Qwen Team. 2024. Qwen2.5 Technical Report.arXiv preprint arXiv:2412.15115 (2024)

2024 arXiv
[34]

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models.arXiv preprint arXiv:2302.13971(2023)

2023 arXiv
[35]

Petar Veličković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph Attention Networks. InInternational Conference on Learning Representations

2018
[36]

Duo Wang, Yuan Zuo, Fengzhi Li, and Junjie Wu. 2024. Llms as zero-shot graph learners: Alignment of gnn representations with llm token embeddings.Advances in neural information processing systems37 (2024), 5950–5973

2024
[37]

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 reason- ing in large language models.Advances in neural information processing systems 35 (2022), 24824–24837

2022
[38]

Zhilin Yang, William Cohen, and Ruslan Salakhudinov. 2016. Revisiting semi- supervised learning with graph embeddings. InInternational conference on ma- chine learning. PMLR, 40–48

2016
[39]

Ruosong Ye, Caiqi Zhang, Runhui Wang, Shuyuan Xu, and Yongfeng Zhang
[40]

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

2024
[41]

Association for Computational Linguistics (ACL), 1955–1973

1955
[42]

recently got into japanese knives and this 120mm wa-petty is my favorite

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). Zhuoyi Peng and Yi Yang A LIMITATIONS AND FUTURE WORK The node text features come from a...

2023 arXiv
[43]

Qihai Zhang, Xinyue Sheng, Yuanfu Sun, and Qiaoyu Tan. 2025. Trustglm: Eval- uating the robustness of graphllms against prompt, text, and structure attacks. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 5912–5923

2025
[45]

anchor_text echo the paper’s own title and abstract
[46]

cluster_token pooled community token
[47]

three_hop_token pooled 3-hop neighbourhood token
[48]

deep learning for image classification with multiple labels

cluster_token re-encode the community predicted label: cs.CV (computer vision, not signal processing) Stage-1 reader decodes this token as: “deep learning for image classification with multiple labels” Stage-1 reader decodes this token as: “deep learning for image captioning” ...
[2024]

In18th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2024-Findings of EACL

Language is All a Graph Needs. In18th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2024-Findings of EACL

2024

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Mohammad Gheshlaghi Azar, Zhaohan Daniel Guo, Bilal Piot, Remi Munos, Mark Rowland, Michal Valko, and Daniele Calandriello. 2024. A general theoret- ical paradigm to understand learning from human preferences. InInternational Conference on Artificial Intelligence and Statistics. PMLR, 4447–4455

2024

[2] [2]

Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. 2024. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 17682–17690

2024

[3] [3]

Runjin Chen, Tong Zhao, Ajay Jaiswal, Neil Shah, and Zhangyang Wang. 2024. LLaGA: large language and graph assistant. InProceedings of the 41st International Conference on Machine Learning. 7809–7823

work page 2024

[4] [4]

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 Newsletter25, 2 (2024), 42–61

work page 2024

[5] [5]

Zhikai Chen, Haitao Mao, Hongzhi Wen, Haoyu Han, Wei Jin, Haiyang Zhang, Hui Liu, and Jiliang Tang. 2024. Label-free node classification on graphs with large language models (llms). InInternational Conference on Learning Representa- tions, Vol. 2024. 31632–31655

work page 2024

[6] [6]

Yuntian Deng, Yejin Choi, and Stuart Shieber. 2024. From explicit cot to implicit cot: Learning to internalize cot step by step.arXiv preprint arXiv:2405.14838 (2024)

arXiv 2024

[7] [7]

Yi Fang, Dongzhe Fan, Daochen Zha, and Qiaoyu Tan. 2024. Gaugllm: Improving graph contrastive learning for text-attributed graphs with large language models. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 747–758

work page 2024

[8] [8]

Sachin Goyal, Ziwei Ji, Ankit Singh Rawat, Aditya Krishna Menon, Sanjiv Kumar, and Vaishnavh Nagarajan. [n. d.]. Think before you speak: Training Language Models With Pause Tokens. InThe Twelfth International Conference on Learning Representations

work page