pith. sign in

arxiv: 2605.24410 · v1 · pith:ZVLIRUOYnew · submitted 2026-05-23 · 💻 cs.AI

Advancing Graph Few-Shot Learning via In-Context Learning

Renchu Guan , Yajun Wang , Chunli Guo , Bowen Cao , Fausto Giunchiglia , Wei Pang , Yonghao Liu , Xiaoyue Feng This is my paper

Pith reviewed 2026-06-30 13:50 UTC · model grok-4.3

classification 💻 cs.AI
keywords graph few-shot learningin-context learningunsupervised meta-learningdual-context fusionnode classificationcontext-aware networkpseudo-tasksfine-tuning-free inference
0
0 comments X

The pith

VISION reframes graph few-shot learning as fine-tuning-free sequence reasoning that fuses local topology with task context in one forward pass.

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

The paper seeks to fix two problems in graph few-shot learning: existing methods ignore abundant unlabeled nodes because they stay inside supervised tasks, and they need extra adaptation steps at inference time. It introduces the VISION model that casts the task as in-context sequence reasoning, much like large language models, so that a single forward pass produces class-aware query representations directly from a support-set context. A context-aware network starts nodes with role embeddings and runs them through a dual-context fusion module; an unsupervised task generator builds pseudo-tasks from unlabeled data to train the whole system. If this works, few-shot node classification becomes practical on large graphs without per-task retraining.

Core claim

VISION reframes graph few-shot learning as a fine-tuning-free sequence reasoning problem. At its core is a context-aware network that initializes nodes with role embeddings and employs a dual-context fusion module to synergistically integrate local topological structures and global task-level dependencies. This allows the model to dynamically generate class-aware representations for the query set conditioned on the support set context in a single forward pass. An unsupervised task generator creates structure-adaptive features and constructs diverse pseudo-tasks from abundant unlabeled data, unifying unsupervised meta-learning with graph in-context learning.

What carries the argument

Dual-context fusion module inside a context-aware network that merges local topological structures with global task-level dependencies to produce class-aware query representations from support-set context.

If this is right

  • Inference requires only a single forward pass with no task-specific adaptation or fine-tuning.
  • Unlabeled nodes are turned into training signal through an unsupervised task generator that builds structure-adaptive pseudo-tasks.
  • The same trained model handles multiple novel classes across different graphs without retraining.
  • Local node topology and global task context are combined inside one module to condition query representations on support examples.

Where Pith is reading between the lines

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

  • The single-pass design could support online updates on streaming graphs where new nodes arrive continuously.
  • Role embeddings that encode support versus query status might transfer to few-shot settings on non-graph structured data such as sequences or tables.
  • Unsupervised pseudo-task construction opens the possibility of scaling the approach to graphs with millions of nodes by sampling diverse subgraphs.

Load-bearing premise

The dual-context fusion module can synergistically integrate local topological structures and global task-level dependencies to dynamically generate accurate class-aware representations for the query set conditioned on the support set context in a single forward pass without any task adaptation or fine-tuning.

What would settle it

Ablation experiments on standard benchmarks where removing the dual-context fusion module produces accuracy no better than fine-tuned baselines, or full-model tests on held-out graphs where accuracy fails to exceed prior methods that require adaptation.

Figures

Figures reproduced from arXiv: 2605.24410 by Bowen Cao, Chunli Guo, Fausto Giunchiglia, Renchu Guan, Wei Pang, Xiaoyue Feng, Yajun Wang, Yonghao Liu.

Figure 1
Figure 1. Figure 1: The overall framework of VISION. Left: The unsupervised meta-training module generates adaptive features [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Sensitivity analysis of model performance. The left panel shows the impact of the number of sampled neighbors [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
read the original abstract

Graph few-shot learning, which aims to classify nodes from novel classes with only a few labeled examples, is a widely studied problem in graph learning. However, existing methods often face two key limitations. First, the predominant graph few-shot learning paradigm relies on supervised tasks, failing to leverage the vast number of unlabeled nodes in the graph. Second, many approaches require complex task adaptation or fine-tuning during inference, limiting their efficiency and applicability. Inspired by the powerful in-context learning capabilities of large language models, we propose a novel model named VISION for adVancIng graph few-Shot learning via In-cOntext LearNing to address these challenges. Our model reframes graph few-shot learning as a fine-tuning-free sequence reasoning problem. At its core is a context-aware network that initializes nodes with role embeddings and employs a dual-context fusion module to synergistically integrate local topological structures and global task-level dependencies. This allows our model to dynamically generate class-aware representations for the query set conditioned on the support set context in a single forward pass. To effectively train our model, we introduce an unsupervised task generator that creates structure-adaptive features and constructs diverse pseudo-tasks from abundant unlabeled data. Our method unifies unsupervised meta-learning with graph in-context learning, achieving efficient inference. Extensive experiments on multiple benchmark datasets demonstrate the superiority of our model. Our public code can be found

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 0 minor

Summary. The manuscript proposes VISION, a model that reframes graph few-shot learning as fine-tuning-free in-context learning. It introduces a context-aware network that uses role embeddings and a dual-context fusion module to integrate local topological structures with global task-level dependencies, enabling single-forward-pass generation of class-aware query representations conditioned on support-set context. Training relies on an unsupervised task generator that constructs pseudo-tasks from unlabeled nodes. The work claims this unifies unsupervised meta-learning with graph in-context learning and demonstrates superiority via extensive experiments on benchmark datasets.

Significance. If the dual-context fusion module and unsupervised task generator function as described, the approach could provide an efficiency advantage over adaptation-based meta-learning methods in graph few-shot settings by enabling inference without task-specific fine-tuning and by exploiting abundant unlabeled graph data.

major comments (2)
  1. [Abstract] Abstract: The central performance claim rests on the dual-context fusion module (with role embeddings) achieving synergistic integration of local and global dependencies to produce accurate class-aware query representations conditioned solely on support-set context in one forward pass. No equations, pseudocode, or architectural diagram specifying the fusion operation (e.g., how support information propagates to queries without leakage or collapse) are supplied, leaving the load-bearing mechanism unverified.
  2. [Abstract] Abstract: The superiority claim and the effectiveness of both the fusion module and the unsupervised task generator are asserted via 'extensive experiments on multiple benchmark datasets,' yet no results, tables, error bars, dataset statistics, ablation studies, or baseline comparisons appear in the manuscript.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their review and the opportunity to address these points. We respond to each major comment below and commit to revisions that improve technical clarity without altering the core claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central performance claim rests on the dual-context fusion module (with role embeddings) achieving synergistic integration of local and global dependencies to produce accurate class-aware query representations conditioned solely on support-set context in one forward pass. No equations, pseudocode, or architectural diagram specifying the fusion operation (e.g., how support information propagates to queries without leakage or collapse) are supplied, leaving the load-bearing mechanism unverified.

    Authors: We agree the abstract is high-level and does not contain the requested technical specifications. The manuscript will be revised to include the equations defining role embeddings and the dual-context fusion operation (cross-attention between support and query with explicit masking to avoid leakage), along with pseudocode and an architectural diagram in Section 3. revision: yes

  2. Referee: [Abstract] Abstract: The superiority claim and the effectiveness of both the fusion module and the unsupervised task generator are asserted via 'extensive experiments on multiple benchmark datasets,' yet no results, tables, error bars, dataset statistics, ablation studies, or baseline comparisons appear in the manuscript.

    Authors: We agree that the current manuscript text does not include the experimental results, tables, or ablations referenced in the abstract. We will add a dedicated experimental section (Section 4) containing these elements, including dataset statistics, baseline comparisons, ablation studies on the fusion module and task generator, and results with error bars on the benchmark datasets. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper proposes a new model VISION that reframes graph few-shot learning as a fine-tuning-free sequence reasoning task, introducing a context-aware network with role embeddings and a dual-context fusion module, plus an unsupervised task generator from unlabeled data. No equations, definitions, or claims in the provided text reduce any central result (such as class-aware query representations or unification of meta-learning with in-context learning) to a fitted input or self-referential construction by definition. The performance claims rest on experimental validation rather than tautological derivation, making the method self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The approach rests on standard domain assumptions in graph machine learning and introduces a new model without additional free parameters or invented physical entities described in the abstract.

axioms (1)
  • domain assumption Graph data contains local topological structures and global task-level dependencies that can be fused to produce useful node representations.
    Invoked when describing the dual-context fusion module that integrates these two sources of information.
invented entities (1)
  • VISION model with context-aware network and dual-context fusion module no independent evidence
    purpose: To enable fine-tuning-free graph few-shot learning via in-context reasoning
    New architecture proposed in the paper; no independent evidence outside the work itself is mentioned.

pith-pipeline@v0.9.1-grok · 5796 in / 1438 out tokens · 29800 ms · 2026-06-30T13:50:01.686673+00:00 · methodology

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

54 extracted references · 4 canonical work pages · 3 internal anchors

  1. [1]

    Amir Bar, Yossi Gandelsman, Trevor Darrell, Amir Globerson, and Alexei Efros

  2. [2]

    InNeurIPS

    Visual prompting via image inpainting. InNeurIPS. 25005–25017

  3. [3]

    Peter L Bartlett and Shahar Mendelson. 2002. Rademacher and gaussian complex- ities: Risk bounds and structural results.Journal of Machine Learning Research3, Nov (2002), 463–482

    2002

  4. [4]

    Aleksandar Bojchevski and Stephan Günnemann. 2018. Deep gaussian embedding of graphs: Unsupervised inductive learning via ranking. InICLR

    2018

  5. [5]

    Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert R Sabuncu, John Guttag, and Adrian V Dalca. 2023. Universeg: Universal medical image segmen- tation. InICCV. 21438–21451

    2023

  6. [6]

    Stephanie Chan, Adam Santoro, Andrew Lampinen, Jane Wang, Aaditya Singh, Pierre Richemond, James McClelland, and Felix Hill. 2022. Data distributional properties drive emergent in-context learning in transformers. InNeurIPS. 18878– 18891

    2022

  7. [7]

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. InNeurIPS. 3837–3845

    2016

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL. 4171–4186

    2019

  9. [9]

    Kaize Ding, Jianling Wang, Jundong Li, Kai Shu, Chenghao Liu, and Huan Liu

  10. [10]

    Graph prototypical networks for few-shot learning on attributed networks. InCIKM. 295–304

  11. [11]

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu, Baobao Chang, Xu Sun, Lei Li, and Zhifang Sui. 2024. A Survey on In-context Learning. InEMNLP. 1107–1128

    2024

  12. [12]

    Junkins, Ehsan Amid, Jure Leskovec, Christopher Ré, and Sebastian Thrun

    Christopher Fifty, Dennis Duan, Ronald G. Junkins, Ehsan Amid, Jure Leskovec, Christopher Ré, and Sebastian Thrun. 2024. Context-Aware Meta-Learning. In ICLR

    2024

  13. [13]

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. Model-agnostic meta- learning for fast adaptation of deep networks. InICML. 1126–1135

    2017

  14. [14]

    Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable feature learning for networks. InKDD. 855–864

    2016

  15. [15]

    Will Hamilton, Zhitao Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. InNeurIPS. 1024–1034

    2017

  16. [16]

    Mikael Henaff, Joan Bruna, and Yann LeCun. 2015. Deep convolutional networks on graph-structured data.arXiv preprint arXiv:1506.05163(2015)

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  17. [17]

    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. InNeurIPS. 22118–22133

    2020

  18. [18]

    Kexin Huang and Marinka Zitnik. 2020. Graph Meta Learning via Local Subgraphs. InNeurIPS. 5862–5874

    2020

  19. [19]

    Kexin Huang and Marinka Zitnik. 2020. Graph meta learning via local subgraphs. InNeurIPS. 5862–5874

    2020

  20. [20]

    Qian Huang, Hongyu Ren, Peng Chen, Gregor Kržmanc, Daniel Zeng, Percy S Liang, and Jure Leskovec. 2023. Prodigy: Enabling in-context learning over graphs. InNeurIPS. 16302–16317

    2023

  21. [21]

    Siavash Khodadadeh, Ladislau Bölöni, and Mubarak Shah. 2019. Unsupervised Meta-Learning for Few-Shot Image Classification. InNeurIPS. 10132–10142

    2019

  22. [22]

    Sungwon Kim, Junseok Lee, Namkyeong Lee, Wonjoong Kim, Seungyoon Choi, and Chanyoung Park. 2023. Task-equivariant graph few-shot learning. InKDD. 1120–1131

    2023

  23. [23]

    Kipf and Max Welling

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

    2017

  24. [24]

    Louis Kirsch, James Harrison, Jascha Sohl-Dickstein, and Luke Metz. 2022. General-purpose in-context learning by meta-learning transformers.arXiv preprint arXiv:2212.04458(2022)

    work page arXiv 2022

  25. [25]

    Yonghao Liu, Lan Huang, Bowen Cao, Ximing Li, Fausto Giunchiglia, Xiaoyue Feng, and Renchu Guan. 2024. A simple but effective approach for unsupervised few-shot graph classification. InWWW. 4249–4259

    2024

  26. [26]

    Yonghao Liu, Mengyu Li, Fausto Giunchiglia, Lan Huang, Ximing Li, Xiaoyue Feng, and Renchu Guan. 2025. Dual-level Mixup for Graph Few-shot Learning with Fewer Tasks. InWWW. 2646–2656

    2025

  27. [27]

    Yonghao Liu, Mengyu Li, Ximing Li, Fausto Giunchiglia, Xiaoyue Feng, and Renchu Guan. 2022. Few-shot node classification on attributed networks with graph meta-learning. InSIGIR. 471–481

    2022

  28. [28]

    Yonghao Liu, Mengyu Li, Ximing Li, Lan Huang, Fausto Giunchiglia, Yanchun Liang, Xiaoyue Feng, and Renchu Guan. 2024. Meta-GPS++: Enhancing graph meta-learning with contrastive learning and self-training.ACM Transactions on Knowledge Discovery from Data18, 9 (2024), 1–30

    2024

  29. [29]

    Yonghao Liu, Chuyao Wang, Zhikang Wang, Liang Chen, Zhi Li, Jiangning Song, Qi Zou, Rui Gao, Binzhi Qian, Xiaoyue Feng, et al. 2026. High-Parameter Spatial Multi-Omics through Histology-Anchored Integration.Nature Methods23, 2 (2026), 373–386

    2026

  30. [30]

    Yonghao Liu, Yajun Wang, Chunli Guo, Wei Pang, Ximing Li, Fausto Giunchiglia, Xiaoyue Feng, and Renchu Guan. 2026. Graph few-shot learning via adaptive spectrum experts and cross-set distribution calibration. InAdvances in Neural Information Processing Systems. 160490–160517. KDD 2026, August 9–13, 2026, Jeju Island, Republic of Korea. Renchu Guan et al

    2026

  31. [31]

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

    2023

  32. [32]

    Ilya Loshchilov and Frank Hutter. 2019. Decoupled Weight Decay Regularization. InICLR

    2019

  33. [33]

    Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. Deepwalk: Online learning of social representations. InKDD. 701–710

    2014

  34. [34]

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. 2019. Language models are unsupervised multitask learners.OpenAI blog 1, 8 (2019), 9

    2019

  35. [35]

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

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  36. [36]

    Aaditya Singh, Stephanie Chan, Ted Moskovitz, Erin Grant, Andrew Saxe, and Felix Hill. 2023. The transient nature of emergent in-context learning in trans- formers. InNeurIPS. 27801–27819

    2023

  37. [37]

    Jake Snell, Kevin Swersky, and Richard Zemel. 2017. Prototypical networks for few-shot learning. InNeurIPS. 4077–4087

    2017

  38. [38]

    Zhen Tan, Ruocheng Guo, Kaize Ding, and Huan Liu. 2023. Virtual node tuning for few-shot node classification. InKDD. 2177–2188

    2023

  39. [39]

    Zhen Tan, Song Wang, Kaize Ding, Jundong Li, and Huan Liu. 2022. Transductive linear probing: A novel framework for few-shot node classification. InLoG

    2022

  40. [40]

    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)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  41. [41]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, Yoshua Bengio, et al. 2018. Graph attention networks. InICLR

    2018

  42. [42]

    Song Wang, Chen Chen, and Jundong Li. 2022. Graph Few-shot Learning with Task-specific Structures. InNeurIPS. 38925–38936

    2022

  43. [43]

    Song Wang, Kaize Ding, Chuxu Zhang, Chen Chen, and Jundong Li. 2022. Task- adaptive few-shot node classification. InKDD. 1910–1919

    2022

  44. [44]

    Song Wang, Yushun Dong, Kaize Ding, Chen Chen, and Jundong Li. 2023. Few- shot node classification with extremely weak supervision. InWSDM. 276–284

    2023

  45. [45]

    Song Wang, Zhen Tan, Huan Liu, and Jundong Li. 2023. Contrastive meta-learning for few-shot node classification. InKDD. 2386–2397

    2023

  46. [46]

    Felix Wu, Amauri Souza, Tianyi Zhang, Christopher Fifty, Tao Yu, and Kilian Weinberger. 2019. Simplifying graph convolutional networks. InICML. 6861– 6871

    2019

  47. [47]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S Yu. 2020. A comprehensive survey on graph neural networks.IEEE Transactions on Neural Networks and Learning Systems32, 1 (2020), 4–24

    2020

  48. [48]

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

    2019

  49. [49]

    Zhilin Yang, William Cohen, and Ruslan Salakhudinov. 2016. Revisiting semi- supervised learning with graph embeddings. InICML. 40–48

    2016

  50. [50]

    Chuxu Zhang, Kaize Ding, Jundong Li, Xiangliang Zhang, Yanfang Ye, Nitesh V Chawla, and Huan Liu. 2022. Few-shot learning on graphs. InIJCAI. 5662–5669

    2022

  51. [51]

    Qiannan Zhang, Shichao Pei, Yuan Fang, and Xiangliang Zhang. 2025. Unlocking the Potential of Black-box Pre-trained GNNs for Graph Few-shot Learning. In AAAI. 22497–22505

    2025

  52. [52]

    Fan Zhou, Chengtai Cao, Kunpeng Zhang, Goce Trajcevski, Ting Zhong, and Ji Geng. 2019. Meta-gnn: On few-shot node classification in graph meta-learning. InCIKM. 2357–2360

    2019

  53. [53]

    Jie Zhou, Ganqu Cui, Shengding Hu, Zhengyan Zhang, Cheng Yang, Zhiyuan Liu, Lifeng Wang, Changcheng Li, and Maosong Sun. 2020. Graph neural networks: A review of methods and applications.AI Open1 (2020), 57–81

    2020

  54. [54]

    Zhiyuan Zhou, Yueming Yin, Hao Han, Yiping Jia, Jun Hong Koh, Adams Wai-Kin Kong, and Yuguang Mu. 2024. ProAffinity-GNN: a novel approach to structure- based protein–Protein binding affinity prediction via a curated data set and graph neural networks.Journal of Chemical Information and Modeling64, 23 (2024), 8796–8808. A Appendix A.1 Verification of Rando...

    2024