Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Prompt-Driven Continual Graph Learning

T0 review · 3 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A frozen GNN with three learned prompts per task matches or beats joint training on four continual graph learning benchmarks, with memory that does not grow with the graph.

desk verdict Solid prompt-based CGL method that is really task-incremental; the authors need to say so. read the letter →

arxiv 2502.06327 v1 pith:FMBI36DU submitted 2025-02-10 cs.LG cs.AI

Qi Wang , Tianfei Zhou , Ye Yuan , Rui Mao This is my paper · ORCID
classification cs.LGcs.AI
keywords continualgraphlearningpromptneuralnetworkscatastrophicforgettingtask-incrementalmemoryefficiencynodeclassificationpersonalizedprompts
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to show that continual learning on graphs does not require replaying old data: a GNN can be pre-trained once, frozen, and then steered through each new task by a small set of learned prompts. The proposed framework stores no historical nodes, only prompts, which removes the memory bottleneck of replay-based methods and the privacy risk of retaining user data. Its two ingredients are hierarchical prompting, which adapts both node features and subgraph-level representations, and a personalized prompt generator that compresses per-node adaptation into a few shared prompts. On CoraFull, OGB-Arxiv, Reddit, and OGB-Products, the paper reports that three prompts per task match or exceed the joint-training upper bound while keeping average forgetting below 0.5%.

What carries the argument

The load-bearing mechanism is hierarchical prompting combined with a personalized prompt generator. Node-level prompts are added to the raw node features before the first GNN layer; subgraph-level prompts are added to the first-layer representations that already carry topological information. For each node $i$, the generator produces a personalized prompt as a weighted sum of the $k$ shared prompts, $p_i^p = \sum_j \alpha_j p_j$ with $\alpha = \mathrm{Softmax}(Q x_i)$, where $Q$ is decomposed as a low-rank outer product $Q = u \otimes v$ to keep parameter cost at $O(k \cdot d)$. This lets the model tailor prompts to individual nodes without storing one prompt per node, which is what makes the memory footprint flat in graph scale.

What would settle it

Run the same four benchmarks in a class-incremental setup where test nodes carry no task labels, or shuffle nodes across task groups before inference; if PROMPTCGL cannot retrieve a prompt without the task identifier, its accuracy should fall to the level of an unprompted or randomly prompted model, showing that the reported gains depend on oracle task IDs.

Watch

Extended reading notes

Core claim

The paper's central claim is that catastrophic forgetting in continual graph learning can be averted by keeping the GNN backbone frozen and learning, for each task, a small prompt bank consisting of node-level prompts and subgraph-level prompts. At inference the prompts for a task are retrieved by the task identifier and injected into the feature stream, so task-specific knowledge lives in the prompt bank rather than in the network weights or a replay buffer. The authors report that PROMPTCGL reaches the joint-training upper bound on all four datasets with only $k=3$ prompts per task, and slightly exceeds it on OGB-Arxiv, while replay-based baselines such as CaT, SSM, and ER-GNN degrade sharply when their memory budget is matched to this small footprint. The memory complexity is $O(k \cdot d)$ with $k=3$, independent of graph size, achieved by a personalized prompt generator whose query matrix is the outer product of a node-dependent vector and a prompt-dependent vector.

Load-bearing premise

The method assumes the task identity of every test node is known at inference; without a task identifier, no prompt can be selected from the bank, and the reported accuracy and forgetting numbers do not apply.

Editorial extensions

If this is right

  • Replay buffers become unnecessary for task-incremental node classification, removing the $O(N \cdot d)$ memory dependence on graph scale.
  • Storing prompts instead of historical nodes avoids retaining raw graph data, addressing the privacy objection to replay-based continual learning.
  • The frozen-backbone design works across the tested GNN architectures (GCN, GAT, GraphSAGE) with no changes to the underlying model, so stronger backbones can be adopted without redesigning the continual learning scheme.
  • Adding a new task costs only $k$ prompts and two low-order query vectors, so the per-task memory footprint stays constant even as graph scale grows.
  • The reported results suggest the joint-training upper bound can be met and even slightly exceeded with task-specific prompting, not merely approached.

Reading between the lines

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

  • The method as presented solves task-incremental continual graph learning, where every test node carries its task identifier; a class-incremental or boundary-free variant would need a learned prompt-selection mechanism, which the paper does not specify.
  • The $O(k \cdot d)$ memory guarantee is per task: over an unbounded stream of tasks the prompt bank grows linearly with the number of tasks, so the constant-memory claim applies to each task's footprint rather than to the whole lifelong run.
  • A natural stress test is to make task graphs more heterogeneous than the four benchmarks, for example with differing feature dimensions or label shifts, and check whether the fixed prompt set $k=3$ still saturates; the rank-1 query decomposition may become the bottleneck.
  • Because prompts are retrieved by task ID, the framework could be extended to few-shot or zero-shot task detection by learning to infer prompts from unlabeled node statistics, but that extension is not evaluated here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes PROMPTCGL, a prompt-based continual graph learning method that keeps a GNN backbone frozen after pre-training on the first task and learns task-specific node-level and subgraph-level prompts produced by a personalized prompt generator. A prompt bank stores the prompts per task; at inference, prompts are retrieved by task identifier. Experiments on CoraFull, OGB-Arxiv, Reddit, and OGB-Products report average performance (AP) and average forgetting (AF), claiming state-of-the-art results, matching or exceeding joint training with only two or three prompts, and O(k·d) memory consumption independent of graph scale.

Significance. If the claims hold in their intended setting, the paper offers a useful alternative to memory replay for continual graph learning, with clear privacy and memory advantages. The work is strengthened by released code, four public benchmarks, component ablations, robustness checks across GCN/GAT/SAGE backbones, and a concrete per-task memory complexity analysis. However, the task-identity assumption and the constrained replay-baseline budget currently prevent the broad state-of-the-art claim from being fully supported.

major comments (3)
  1. [III-B5 (Eq. 8), Algorithm 1, Section III-A] Inference in Eq. (8) retrieves prompt parameters Pi 'based on the task identifier', and Algorithm 1 stores prompts per task. The problem definition in Section III-A never states that task identity is available at test time, and no limitation section acknowledges this requirement. Consequently, the reported AP/AF values are for task-incremental continual learning with an oracle task identity, not for class-incremental or general CGL. This matters because the low forgetting is partly by construction: prompts of previous tasks are frozen and only the shared prediction layer is updated, so the hardest sub-problem of deciding which task a test node belongs to is removed. Replay baselines such as ER-GNN, SSM, and CaT do not require task IDs at inference, so the comparison is not on equal footing with respect to task identity. Please explicitly scope the method and claims to task-incremental CGL, or evaluate under a setting without task identity and state the limitation.
  2. [Table II, Section IV-E1] The replay baselines are evaluated only under a memory budget equivalent to 2 or 3 prompt vectors, i.e., roughly 3-5 stored nodes per task as stated in the table. This is far below the replay ratios (e.g., krate = 0.01) used in the CGLB benchmark and in the original papers of ER-GNN, SSM, and CaT. The conclusion that PROMPTCGL 'achieves superior performance against existing CGL approaches' is therefore not supported in the general case; the evidence supports only the claim that prompts outperform replay methods under an extremely tight memory constraint. Please add experiments at the standard replay buffer sizes, or at least report the baselines in their standard configurations, and adjust the state-of-the-art claim accordingly.
  3. [Eqs. (5)-(6), Section III-B3] The dimensions in the personalized prompt generator are inconsistent. Eq. (5) requires Q x0_i to be a k-dimensional vector because α = Softmax(...) weights k prompts, so Q must have shape k × d_f. Eq. (6) defines Q = u ⊗ v with u ∈ R^{1×n} and v ∈ R^{1×k}, whose outer product has shape n × k, not k × d_f. If n denotes the number of tasks as in Section III-A, then the low-order vector u also grows with the number of tasks, contradicting the stated O(k·d) memory complexity that is independent of task count. Please correct the notation, give the actual tensor shapes, and verify the memory-complexity claim with these shapes.
minor comments (5)
  1. [III-B2, Eq. (4) text] The sentence defining Xp_1 states Xp_1 ∈ R^{k×dh}, but the prompted representations are per-node and are fed into subsequent GNN layers, so the shape should be R^{N×dh}; the subgraph-level prompt set has size k, but the personalized prompts are generated for every node.
  2. [IV-C, Eq. (9)] The task indexing is inconsistent: the text says tasks are indexed 0,1,...,T−1, while the formulas sum q=1 to T for AP and q=1 to T−1 for AF. Please unify the indexing convention.
  3. [Algorithm 1 and Section III-B1] The algorithm learns prompts for every task in D, including the initial task T0, whereas Section III-B1 says the backbone and prediction layer are pre-trained on T0 without prompts; clarify whether T0 is included in the prompt-learning loop.
  4. [III-B3, Eq. (6)] The symbol n is overloaded: it denotes the number of tasks in Section III-A and is used as a dimension in Eq. (6). Please rename one of the two usages to avoid ambiguity.
  5. [Fig. 6] The axis labels in the running-time figure are truncated (e.g., 'Running Time ( )' and 'AP ( )') and the units are missing; please fix the labels and include units.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's contributions are empirical and architectural; memory claims follow from definitions and accuracy claims are external benchmark comparisons.

full rationale

The paper's central claims are empirical evaluations against external public benchmarks. Prompt parameters are optimized per task via Eq. (7) and evaluated on held-out test partitions in Table II, so the reported AP/AF numbers are not derived from, or fitted to, the claims they are meant to support. The O(k*d) memory statement in Section III-D is an arithmetic consequence of the defined prompt dimensions and the query decomposition in Eqs. (2)-(6); it is not a fitted result renamed as a prediction. Low forgetting is measured through Eq. (9) rather than posited by definition: although each task's prompts are saved in the prompt bank, the shared prediction layer is still updated across tasks, so the small AF values are empirical outcomes rather than constructional equivalences. The paper also does not rely on a load-bearing self-citation or an imported uniqueness theorem; self-references such as [6] are background citations for GNN applications and do not justify the main contribution. The task-identifier requirement in Section III-B5 and Algorithm 1 is a legitimate scoping and baseline-fairness concern, but it concerns the generality of the continual-learning setting, not a circular reduction of the paper's stated predictions to its inputs.

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

The central claims rest on the task-incremental protocol, the transferability of a frozen pretrained backbone under additive prompts, and tuned hyperparameters including prompt count and learning rates. No unobserved physical entities are introduced. The prompts, prompt bank, and personalized prompt generator are learnable model components with direct empirical validation on public benchmarks.

free parameters (3)
  • prompt_count_k = 3
    Ablated over 1, 2, 3, and 4 prompts on CoraFull and Arxiv (Table IV); k=3 is reported as optimal, with k=2 nearly equal. The constant-memory claim depends on k staying small.
  • prompt_learning_rate_alpha = 0.01
    Grid-searched on Arxiv (Fig. 7, right). AP degrades when alpha is too small, so it is a tuned hyperparameter rather than a derived quantity.
  • prediction_layer_learning_rate_beta = 5e-4
    Grid-searched on Arxiv (Fig. 7, left). Set lower than alpha to slow drift of the shared prediction layer, which is a manual tuning choice.
assumptions (3)
  • domain assumption Task identity is known at inference and used to retrieve the correct prompt set.
    Eq. (8) and Algorithm 1 select prompts by task identifier. Without this, the model cannot choose which prompts to apply, so the reported forgetting and accuracy results are contingent on the task-incremental protocol.
  • domain assumption A GNN backbone frozen after task T0 can be adapted to all later task graphs by additive prompts alone.
    The backbone is never updated after pretraining on the first two classes (Section III.B). The whole method relies on prompts shifting features and hidden representations enough for new classes.
  • domain assumption The four public benchmarks and 60/20/20 class-wise splits are a valid operationalization of continual graph learning.
    All results are measured on this protocol from [11] and [20]. The scope of the claims is limited to this protocol.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Prompt-Driven Continual Graph Learning." pith.science (2026). https://pith.science/paper/FMBI36DU

@misc{pith2026250206327,
  author       = {Pith},
  title        = {Pith review of: Prompt-Driven Continual Graph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FMBI36DU}},
  note         = {Machine review of arXiv:2502.06327}
}
read the original abstract

Continual Graph Learning (CGL), which aims to accommodate new tasks over evolving graph data without forgetting prior knowledge, is garnering significant research interest. Mainstream solutions adopt the memory replay-based idea, ie, caching representative data from earlier tasks for retraining the graph model. However, this strategy struggles with scalability issues for constantly evolving graphs and raises concerns regarding data privacy. Inspired by recent advancements in the prompt-based learning paradigm, this paper introduces a novel prompt-driven continual graph learning (PROMPTCGL) framework, which learns a separate prompt for each incoming task and maintains the underlying graph neural network model fixed. In this way, PROMPTCGL naturally avoids catastrophic forgetting of knowledge from previous tasks. More specifically, we propose hierarchical prompting to instruct the model from both feature- and topology-level to fully address the variability of task graphs in dynamic continual learning. Additionally, we develop a personalized prompt generator to generate tailored prompts for each graph node while minimizing the number of prompts needed, leading to constant memory consumption regardless of the graph scale. Extensive experiments on four benchmarks show that PROMPTCGL achieves superior performance against existing CGL approaches while significantly reducing memory consumption. Our code is available at https://github.com/QiWang98/PromptCGL.

Figures

Figures reproduced from arXiv: 2502.06327 by the authors.

Figure 1
Figure 1. Main Idea. Replay-based methods, e.g., CaT [20], SSM [18], ER￾GNN [17], require a memory buffer to store a number of graph nodes per task, which is merged with the incoming graph for model retraining (see (a)). However, they face a severe degradation when the buffer size decreases (see (b) and (c)). In contrast, PROMPTCGL represents a novel prompt￾based learning paradigm, which learns a fixed number of prompts for e… view at source ↗
Figure 2
Figure 2. Illustration of PROMPTCGL framework. Here we present the execution steps for task Tt. All tasks except T0 follow the same procedure. The backbone parameters, pre-trained on task T0, remain frozen in subsequent tasks. Initially, node-level personalized prompts are generated by the personalized prompt generator (PG) based on the query result of the node feature and a maintained small node-level prompt set, which are t… view at source ↗
Figure 3
Figure 3. Illustration of the Personalized Prompt Generator. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Performance matrix visualization of Joint, Ours, CaT, SSM, MAS and GEM on CoraFull, Arxiv, Reddit and Products datasets (from top to bottom). [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The visualization of node embedding learned without (left) and with prompts (right) on four datasets. [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Comparison of running time and AP on CoraFull (left) and Products [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Comparison of different learning rates of prediction layer (left) and [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Task-Aware Adaptive Modulation: A Replay-Free and Resource-Efficient Approach For Continual Graph Learning

    cs.LG 2025-08 conditional novelty 6.0 of 10

    TAAM achieves replay-free, pre-training-free continual graph learning by routing each task to a small node-attentive modulator inserted into a frozen GNN.

Reference graph

Works this paper leans on

52 extracted references · 37 canonical work pages · cited by 1 Pith paper

  1. [1]

    A comprehensive survey on graph neural networks,

    Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and S. Y . Philip, “A comprehensive survey on graph neural networks,” IEEE transactions on neural networks and learning systems , vol. 32, no. 1, pp. 4–24, 2020

    2020

  2. [2]

    Efficient subgraph search over large uncertain graphs,

    Y . Yuan, G. Wang, H. Wang, and L. Chen, “Efficient subgraph search over large uncertain graphs,” Proceedings of the VLDB Endowment , vol. 4, no. 11, pp. 876–886, 2011

    work page 2011

  3. [3]

    Cgmega: explainable graph neural network framework with attention mechanisms for cancer gene module dissection,

    H. Li, Z. Han, Y . Sun, F. Wang, P. Hu, Y . Gao, X. Bai, S. Peng, C. Ren, X. Xu et al. , “Cgmega: explainable graph neural network framework with attention mechanisms for cancer gene module dissection,” Nature Communications, vol. 15, no. 1, p. 5997, 2024

    work page 2024

  4. [4]

    A weighted symmetric graph embedding approach for link prediction in undirected graphs,

    Z. Wang, Y . Chai, C. Sun, X. Rui, H. Mi, X. Zhang, and S. Y . Philip, “A weighted symmetric graph embedding approach for link prediction in undirected graphs,” IEEE Transactions on Cybernetics, vol. 54, no. 2, pp. 1037–1047, 2022

    work page 2022

  5. [5]

    Semisuper- vised graph neural networks for graph classification,

    Y . Xie, Y . Liang, M. Gong, A. K. Qin, Y .-S. Ong, and T. He, “Semisuper- vised graph neural networks for graph classification,” IEEE Transactions on Cybernetics, vol. 53, no. 10, pp. 6222–6235, 2022

    work page 2022

  6. [6]

    Noise-resistant graph neural networks for session-based recommenda- tion,

    Q. Wang, A. Wu, Y . Yuan, Y . Wang, G. Zhong, X. Gao, and C. Yang, “Noise-resistant graph neural networks for session-based recommenda- tion,” in Proceedings APWeb and WAIM Joint International Conference on Web and Big Data . Springer, 2024, pp. 144–160

    work page 2024

  7. [7]

    Grakerformer: A transformer with graph kernel for unsupervised graph representation learning,

    L. Xu, H. Liu, X. Yuan, E. Chen, and Y . Tang, “Grakerformer: A transformer with graph kernel for unsupervised graph representation learning,” IEEE Transactions on Cybernetics , vol. 54, no. 12, pp. 7320– 7332, 2024

    work page 2024

  8. [8]

    Overcoming catastrophic forgetting in graph neural networks,

    H. Liu, Y . Yang, and X. Wang, “Overcoming catastrophic forgetting in graph neural networks,” in Proceedings of the AAAI conference on artificial intelligence, vol. 35, no. 10, 2021, pp. 8653–8661

    work page 2021

Show all 52 references
  1. [9]

    Lifelong graph learning,

    C. Wang, Y . Qiu, D. Gao, and S. Scherer, “Lifelong graph learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 13 719–13 728

    2022

  2. [10]

    Adaptive cointegration analysis and modified rpca with continual learning ability for monitoring multimode nonstationary processes,

    J. Zhang, D. Zhou, and M. Chen, “Adaptive cointegration analysis and modified rpca with continual learning ability for monitoring multimode nonstationary processes,” IEEE Transactions on Cybernetics , vol. 53, no. 8, pp. 4841–4854, 2022

    2022

  3. [11]

    Cglb: Benchmark tasks for continual graph learning,

    X. Zhang, D. Song, and D. Tao, “Cglb: Benchmark tasks for continual graph learning,” Advances in Neural Information Processing Systems , vol. 35, pp. 13 006–13 021, 2022

    2022

  4. [12]

    Rein- forced continual learning for graphs,

    A. Rakaraddi, L. Siew Kei, M. Pratama, and M. De Carvalho, “Rein- forced continual learning for graphs,” in Proceedings of the 31st ACM International Conference on Information & Knowledge Management , 2022, pp. 1666–1674

    2022

  5. [13]

    Multimodal continual graph learning with neural architecture search,

    J. Cai, X. Wang, C. Guan, Y . Tang, J. Xu, B. Zhong, and W. Zhu, “Multimodal continual graph learning with neural architecture search,” in Proceedings of the ACM Web Conference 2022 , 2022, pp. 1292–1300

    2022

  6. [14]

    Continual learning on graphs: Chal- lenges, solutions, and opportunities,

    X. Zhang, D. Song, and D. Tao, “Continual learning on graphs: Chal- lenges, solutions, and opportunities,” arXiv preprint arXiv:2402.11565 , 2024

    2024 arXiv

  7. [15]

    Hierarchical prototype networks for continual graph representation learning,

    X. Zhang, D. Song, and D. Tao, “Hierarchical prototype networks for continual graph representation learning,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 45, no. 4, pp. 4622–4636, 2023

    2023

  8. [16]

    Continual learning on dynamic graphs via parameter isolation,

    P. Zhang, Y . Yan, C. Li, S. Wang, X. Xie, G. Song, and S. Kim, “Continual learning on dynamic graphs via parameter isolation,” in Pro- ceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval , 2023, pp. 601–611

    2023

  9. [17]

    Overcoming catastrophic forgetting in graph neural networks with experience replay,

    F. Zhou and C. Cao, “Overcoming catastrophic forgetting in graph neural networks with experience replay,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 35, no. 5, 2021, pp. 4714–4722

    2021

  10. [18]

    Sparsified subgraph memory for continual graph representation learning,

    X. Zhang, D. Song, and D. Tao, “Sparsified subgraph memory for continual graph representation learning,” in Proceeding of the IEEE International Conference on Data Mining . 2022, pp. 1335–1340

    2022

  11. [19]

    Dygrain: An incremental learning frame- work for dynamic graphs

    S. Kim, S. Yun, and J. Kang, “Dygrain: An incremental learning frame- work for dynamic graphs.” inProceedings of the Thirty-First International Joint Conference on Artificial Intelligence , 2022, pp. 3157–3163

    2022

  12. [20]

    Cat: Balanced continual graph learning with graph condensation,

    Y . Liu, R. Qiu, and Z. Huang, “Cat: Balanced continual graph learning with graph condensation,” in Proceeding of the IEEE International Conference on Data Mining , 2023, pp. 1157–1162

    2023

  13. [21]

    Ricci curvature-based graph sparsifi- cation for continual graph representation learning,

    X. Zhang, D. Song, and D. Tao, “Ricci curvature-based graph sparsifi- cation for continual graph representation learning,” IEEE Transactions on Neural Networks and Learning Systems, vol. 35, no. 12, pp. 17398-17410, 2024

    2024

  14. [22]

    A federated graph neural network framework for privacy-preserving personalization,

    C. Wu, F. Wu, L. Lyu, T. Qi, Y . Huang, and X. Xie, “A federated graph neural network framework for privacy-preserving personalization,”Nature Communications, vol. 13, no. 1, p. 3091, 2022

    2022

  15. [23]

    Making pre-trained language models end-to-end few-shot learners with contrastive prompt tuning,

    Z. Xu, C. Wang, M. Qiu, F. Luo, R. Xu, S. Huang, and J. Huang, “Making pre-trained language models end-to-end few-shot learners with contrastive prompt tuning,” in Proceedings of the Sixteenth ACM Interna- tional Conference on Web Search and Data Mining , 2023, pp. 438–446

    2023

  16. [24]

    Learning to prompt for continual learning,

    Z. Wang, Z. Zhang, C.-Y . Lee, H. Zhang, R. Sun, X. Ren, G. Su, V . Perot, J. Dy, and T. Pfister, “Learning to prompt for continual learning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 139–149. 12

    2022

  17. [25]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing,

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, and G. Neubig, “Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing,” ACM Computing Surveys, vol. 55, no. 9, pp. 1–35, 2023

    2023

  18. [26]

    Prefix-tuning: Optimizing continuous prompts for generation,

    X. L. Li and P. Liang, “Prefix-tuning: Optimizing continuous prompts for generation,” in Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), ...

    2021

  19. [27]

    Universal prompt tuning for graph neural networks,

    T. Fang, Y . Zhang, Y . Yang, C. Wang, and L. Chen, “Universal prompt tuning for graph neural networks,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  20. [28]

    All in one: Multi-task prompting for graph neural networks,

    X. Sun, H. Cheng, J. Li, B. Liu, and J. Guan, “All in one: Multi-task prompting for graph neural networks,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023, pp. 2120–2131

    2023

  21. [29]

    Graph prompt learning: A comprehensive survey and beyond,

    X. Sun, J. Zhang, X. Wu, H. Cheng, Y . Xiong, and J. Li, “Graph prompt learning: A comprehensive survey and beyond,” arXiv preprint arXiv:2311.16534, 2023

    2023 arXiv

  22. [30]

    Virtual node tuning for few-shot node classification,

    Z. Tan, R. Guo, K. Ding, and H. Liu, “Virtual node tuning for few-shot node classification,” in Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2023, pp. 2177–2188

    2023

  23. [31]

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

    M. Sun, K. Zhou, X. He, Y . Wang, and X. Wang, “Gppt: Graph pre-training and prompt tuning to generalize graph neural networks,” in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2022, pp. 1717–1727

    2022

  24. [32]

    It’s not just size that matters: Small language models are also few-shot learners,

    T. Schick and H. Sch ¨utze, “It’s not just size that matters: Small language models are also few-shot learners,” inProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , 2021, pp. 2339–2352

    2021

  25. [33]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  26. [34]

    Learning to prompt for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision-language models,” International Journal of Computer Vision , vol. 130, no. 9, pp. 2337–2348, 2022

    2022

  27. [35]

    Conditional prompt learning for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Conditional prompt learning for vision-language models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16 816–16 825

    2022

  28. [36]

    Visual prompt tuning,

    M. Jia, L. Tang, B.-C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S.-N. Lim, “Visual prompt tuning,” in Proceedings of the European Conference on Computer Vision , 2022, pp. 709–727

    2022

  29. [37]

    Adaptformer: Adapting vision transformers for scalable visual recog- nition,

    S. Chen, C. Ge, Z. Tong, J. Wang, Y . Song, J. Wang, and P. Luo, “Adaptformer: Adapting vision transformers for scalable visual recog- nition,” Advances in Neural Information Processing Systems , vol. 35, pp. 16 664–16 678, 2022

    2022

  30. [38]

    Prompt switch: Efficient clip adaptation for text-video retrieval,

    C. Deng, Q. Chen, P. Qin, D. Chen, and Q. Wu, “Prompt switch: Efficient clip adaptation for text-video retrieval,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 15 648–15 658

    2023

  31. [39]

    Promptcap: Prompt-guided task-aware image captioning,

    Y . Hu, H. Hua, Z. Yang, W. Shi, N. A. Smith, and J. Luo, “Promptcap: Prompt-guided task-aware image captioning,” arXiv preprint arXiv:2211.09699, 2022

    2022 arXiv

  32. [40]

    Exploring low-resource medical image classification with weakly supervised prompt learning,

    F. Zheng, J. Cao, W. Yu, Z. Chen, N. Xiao, and Y . Lu, “Exploring low-resource medical image classification with weakly supervised prompt learning,” Pattern Recognition, vol. 149, p. 110250, 2024

    2024

  33. [41]

    Learning to prompt for open-vocabulary object detection with vision-language model,

    Y . Du, F. Wei, Z. Zhang, M. Shi, Y . Gao, and G. Li, “Learning to prompt for open-vocabulary object detection with vision-language model,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 14 084–14 093

    2022

  34. [42]

    Learning background prompts to discover implicit knowledge for open vocabu- lary object detection,

    J. Li, J. Zhang, J. Li, G. Li, S. Liu, L. Lin, and G. Li, “Learning background prompts to discover implicit knowledge for open vocabu- lary object detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 16 678–16 687

    2024

  35. [43]

    Vscode: General visual salient and camouflaged object detection with 2d prompt learning,

    Z. Luo, N. Liu, W. Zhao, X. Yang, D. Zhang, D.-P. Fan, F. Khan, and J. Han, “Vscode: General visual salient and camouflaged object detection with 2d prompt learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 17 169–17 180

    2024

  36. [44]

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

    Z. Liu, X. Yu, Y . Fang, and X. Zhang, “Graphprompt: Unifying pre- training and downstream tasks for graph neural networks,” in Proceedings of the ACM Web Conference 2023 , 2023, pp. 417–428

    2023

  37. [45]

    Automating the construction of internet portals with machine learning,

    A. K. McCallum, K. Nigam, J. Rennie, and K. Seymore, “Automating the construction of internet portals with machine learning,” Information Retrieval, vol. 3, pp. 127–163, 2000

    2000

  38. [46]

    Open graph benchmark: Datasets for machine learning on graphs,

    W. Hu, M. Fey, M. Zitnik, Y . Dong, H. Ren, B. Liu, M. Catasta, and J. Leskovec, “Open graph benchmark: Datasets for machine learning on graphs,” Advances in neural information processing systems , vol. 33, pp. 22 118–22 133, 2020

    2020

  39. [47]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Advances in neural information processing systems, vol. 30, 2017

    2017

  40. [48]

    Overcoming catastrophic forgetting in neural networks,

    J. Kirkpatrick, R. Pascanu, N. Rabinowitz, J. Veness, G. Desjardins, A. A. Rusu, K. Milan, J. Quan, T. Ramalho, A. Grabska-Barwinska et al., “Overcoming catastrophic forgetting in neural networks,” Proceedings of the national academy of sciences , vol. 114, no. 13, pp. 3521–3526, 2017

    2017

  41. [49]

    Memory aware synapses: Learning what (not) to forget,

    R. Aljundi, F. Babiloni, M. Elhoseiny, M. Rohrbach, and T. Tuytelaars, “Memory aware synapses: Learning what (not) to forget,” in Proceedings of the European conference on computer vision , 2018, pp. 139–154

    2018

  42. [50]

    Gradient episodic memory for continual learning,

    D. Lopez-Paz and M. Ranzato, “Gradient episodic memory for continual learning,” Advances in neural information processing systems , vol. 30, 2017

    2017

  43. [51]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,” IEEE transactions on pattern analysis and machine intelligence , vol. 40, no. 12, pp. 2935– 2947, 2017

    2017

  44. [52]

    Graph attention networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Lio, and Y . Bengio, “Graph attention networks,”arXiv preprint arXiv:1710.10903, 2017

    2017 arXiv

Pith tools

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