arxiv: 2605.12061 · v1 · submitted 2026-05-12 · 💻 cs.AI

Recognition: 1 theorem link

· Lean Theorem

SAGE: A Self-Evolving Agentic Graph-Memory Engine for Structure-Aware Associative Memory

Juntong Wang , Haoyue Zhao , guanghui Pan , Xiyuan Wang , Yanbo Wang , Qiyan Deng , Muhan Zhang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 04:39 UTC · model grok-4.3

classification 💻 cs.AI

keywords graph memoryself-evolving agentslong-term memorymulti-hop QAhallucination reductionagentic systemsmemory writermemory reader

0 comments

The pith

SAGE lets language agents evolve a graph memory by having a writer add structure from interactions and a reader supply feedback to refine it.

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

The paper argues that static retrieval graphs limit agents because they cannot recover full evidence chains from partial cues or improve themselves from downstream use. SAGE addresses this by pairing a memory writer that incrementally builds the graph with a graph foundation model reader that retrieves answers and sends corrective signals back to the writer. Over repeated rounds this closed loop produces measurable gains in evidence recovery, answer grounding, and hallucination reduction. The approach is tested on multi-hop question answering, open-domain retrieval, and long-term agent memory benchmarks, where two evolution rounds yield the highest average rank and strong zero-shot transfer results.

Core claim

The central claim is that modeling long-term memory as a dynamic graph substrate, with incremental construction by the writer and diagnostic feedback from the reader, creates genuine self-evolution that improves the agent's ability to link partial cues to complete evidence chains and ground responses more reliably.

What carries the argument

The writer-reader feedback loop, in which the memory writer incrementally constructs structured graph memory from interaction histories while the graph foundation model reader performs retrieval and returns refinement signals.

If this is right

After two self-evolution rounds the system achieves the best average rank across multi-hop QA tasks.
Zero-shot transfer to open-domain retrieval reaches 82.5 recall at rank 2 and 91.6 at rank 5 on Natural Questions.
Reader-writer feedback during training improves multiple metrics on long-term memory and hallucination diagnostics.
The evolved graph supports more complete evidence recovery and better answer grounding than static retrieval baselines.

Where Pith is reading between the lines

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

The same loop could let agents accumulate reusable structural roles across many tasks without explicit retraining.
Extending the reader feedback to other agent modules might further reduce reliance on large context windows.
Comparing graph states before and after evolution rounds could identify which added edges most reduce hallucinations.
The method might apply to non-text domains if analogous writer and reader roles are defined for other data types.

Load-bearing premise

The assumption that incremental graph construction plus reader feedback produces stable self-evolution that improves evidence recovery without instability or overfitting to the benchmarks.

What would settle it

Running the full writer-reader loop for multiple rounds on a new multi-hop QA dataset never seen during development and observing no increase in recall or no reduction in hallucination rates.

Figures

Figures reproduced from arXiv: 2605.12061 by guanghui Pan, Haoyue Zhao, Juntong Wang, Muhan Zhang, Qiyan Deng, Xiyuan Wang, Yanbo Wang.

**Figure 2.** Figure 2: Overall pipeline of the proposed SAGE. The memory writer incrementally constructs and updates graph memory from observations through state-conditioned writing actions, and receives rewards from downstream memory use. The resulting retrieval feedback closes the loop between writing and reading, enabling graph memory to improve over time. Graph Foundation Models. Graph Foundation Models (GFMs) aim to learn t… view at source ↗

**Figure 3.** Figure 3: Visualization of the retrieved results. category comprises long-term agent memory datasets, including LongMemEval and HaluMem, used to test information extraction from long interaction histories, multi-session reasoning, temporal reasoning, knowledge updating, abstention, and operation-level hallucination [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

**Figure 4.** Figure 4: Freeze the sensitivity analysis on the reader side. The impact of the initial-entity weight [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗

**Figure 5.** Figure 5: Training regularization and scaling analysis. A larger rollout group size and a moderate [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗

**Figure 6.** Figure 6: Metadata of the case study from HotpotQA. Inferer prompt template. You are a retrieval planner for graph-based multi-hop QA. Question: {QUESTION} Structured extraction: {EXTRACTOR_JSON} Generate at most M retrieval intents that help locate: - evidence directly supporting the target relation; - bridge entities required for multi-hop reasoning; - documents likely to contain the target attribute; - evidence s… view at source ↗

**Figure 7.** Figure 7: Metadata of the case study from HotpotQA. It consists of two stages: Extractor 6 extracts explicit entities, aliases, relation clues, hard constraints, and the answer type; Inferer 7 generates at most M retrieval intents based on the extraction results. L Computation Details of Topological Structural Features L.1 Normalized Structural Graph Structural features are computed on an undirected, self-loop-free,… view at source ↗

read the original abstract

Long-term memory is becoming a central bottleneck for language agents. Exsting RAG and GraphRAG systems largely treat memory graphs as static retrieval middleware, which limits their ability to recover complete evidence chains from partial cues, exploit reusable graph-structrual roles, and improve the memory itself through downstream feedback. We introduce SAGE, a Self-evolving Agentic Graph-memory Engine that models graph memory as a dynamic long-term memory substrate. SAGE couples two roles: a memory writer that incrementally constucts structured graph memory from interaction histories, and a Graph Foundation Model-based memory reader to perform retrieval and provide feedback to the memory writer. We provide rigorooous theoretical annalyses supporting the framework. Across multi-hop QA, open-domain retireval, domain-specific review QA, and long-term agent-memory benchmarks, SAGE improves evidence recovery, answer grounding, and retrieval efficiency: after two self-evolution rounds, it achieves the best average rank on multi-hop QA; in zero-shot open-domain transfer, it reaches 82.5/91.6 Recall@2/5 on NQ. Further results on LongMemEval and HaluMem show that traning and reader-writer feedback improve multiple long-term memory and hallucination-diagnostic metrics, suggesting that self-evolving, structure-aware graph memory is a promising foundation for robust long-horizon language agents.

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.

Desk Editor's Note private letter to a colleague

SAGE adds a writer-reader feedback loop to make graph memory self-evolving for agents and claims clear benchmark gains, but lacks ablations to show the evolution itself drives the improvements.

read the letter

The main point with this SAGE paper is that it proposes a self-evolving agentic graph memory where a writer incrementally constructs the graph from interaction histories and a Graph Foundation Model reader retrieves information while providing feedback to refine the memory. After two self-evolution rounds, it claims the best average rank on multi-hop QA and reaches 82.5/91.6 Recall@2/5 on NQ in zero-shot open-domain transfer, with additional gains on long-term memory and hallucination benchmarks. What is new here is the explicit writer-reader loop that allows the memory to improve itself through feedback, moving past static graph retrieval systems. The paper does a solid job identifying the limitations of current RAG and GraphRAG approaches for recovering complete evidence chains and exploiting graph structures in agent settings. The results on evidence recovery and reduced hallucinations from the training and feedback are encouraging if they replicate. However, the soft spots include the absence of detailed ablations that would show whether the gains come specifically from the self-evolution mechanism or simply from additional iterations or data exposure. The assumption that incremental graph updates plus reader feedback lead to genuine structural improvements without instability needs more support. The abstract mentions rigorous theoretical analyses, but without seeing the actual content, it's difficult to assess their rigor. A few typos in the provided abstract also suggest some lack of polish. This paper targets researchers in AI agents and memory systems. Someone looking for ideas on dynamic long-term memory would find the framework useful to consider. I would recommend putting it through peer review. The problem it tackles is timely, the architecture is clearly described at a high level, and the empirical claims are specific enough that referees can evaluate the methods and results properly.

Referee Report

2 major / 2 minor

Summary. The paper introduces SAGE, a Self-evolving Agentic Graph-memory Engine that treats graph memory as a dynamic substrate for language agents. It couples a memory writer that incrementally constructs structured graphs from interaction histories with a Graph Foundation Model-based memory reader that performs retrieval and supplies feedback to the writer. The framework is supported by claimed rigorous theoretical analyses. Empirically, after two self-evolution rounds SAGE attains the best average rank across multi-hop QA tasks, achieves 82.5/91.6 Recall@2/5 on NQ in zero-shot open-domain transfer, and improves long-term memory and hallucination-diagnostic metrics on LongMemEval and HaluMem.

Significance. If the self-evolution mechanism produces stable structural improvements in evidence recovery independent of extra computation or benchmark-specific patterns, the work would meaningfully advance dynamic, structure-aware memory beyond static RAG and GraphRAG systems. The reported gains on multi-hop QA ranking, NQ transfer, and hallucination metrics on LongMemEval/HaluMem constitute a concrete empirical contribution that could inform more reliable long-horizon agents. The explicit coupling of writer construction with reader feedback is a clear architectural strength.

major comments (2)

[§5 (Experiments)] §5 (Experiments): The central claim that two self-evolution rounds yield the best average rank on multi-hop QA and the 82.5/91.6 R@2/5 on NQ rests on the assumption that reader-writer feedback produces genuine structural improvement. No ablation isolating the contribution of the evolution rounds from additional writer-reader iterations, extra interaction histories, or increased compute is reported, leaving open the possibility that gains are artifacts of extra computation rather than self-evolution.
[§3 (Theoretical Analyses)] §3 (Theoretical Analyses): The abstract asserts 'rigorous theoretical analyses supporting the framework,' yet no equations, lemmas, convergence proofs, or formal arguments appear that demonstrate stability of incremental graph updates, non-monotonicity bounds, or guarantees against overfitting to benchmark patterns. This is load-bearing because the self-evolution claim requires such analysis to distinguish it from simple iterative refinement.

minor comments (2)

[Abstract] Abstract: Multiple typographical errors ('Exsting', 'constucts', 'rigorooous theoretical annalyses', 'retireval', 'traning') should be corrected for clarity.
[Abstract and §2] Abstract and §2: The term 'Graph Foundation Model' is used without definition or citation on first appearance.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments identify key areas where our claims on self-evolution and theoretical grounding can be strengthened. We address each major comment below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

Referee: §5 (Experiments): The central claim that two self-evolution rounds yield the best average rank on multi-hop QA and the 82.5/91.6 R@2/5 on NQ rests on the assumption that reader-writer feedback produces genuine structural improvement. No ablation isolating the contribution of the evolution rounds from additional writer-reader iterations, extra interaction histories, or increased compute is reported, leaving open the possibility that gains are artifacts of extra computation rather than self-evolution.

Authors: We agree that the absence of targeted ablations leaves the self-evolution claim open to the interpretation that gains arise from additional iterations or compute rather than structural refinement. In the revised manuscript we will add ablations that hold total writer-reader interactions and compute fixed while varying only the presence of feedback-driven graph updates (e.g., comparing self-evolution against frozen graphs or non-adaptive edge additions over the same number of rounds). These results will be reported alongside the existing multi-hop QA and NQ numbers to isolate the contribution of the evolving structure. revision: yes
Referee: §3 (Theoretical Analyses): The abstract asserts 'rigorous theoretical analyses supporting the framework,' yet no equations, lemmas, convergence proofs, or formal arguments appear that demonstrate stability of incremental graph updates, non-monotonicity bounds, or guarantees against overfitting to benchmark patterns. This is load-bearing because the self-evolution claim requires such analysis to distinguish it from simple iterative refinement.

Authors: We acknowledge that §3 currently offers conceptual reasoning on the benefits of incremental graph construction and reader feedback rather than formal lemmas or convergence proofs. The abstract's phrasing therefore overstates the level of rigor. In revision we will (i) replace 'rigorous theoretical analyses' in the abstract with a more precise description of the conceptual arguments provided and (ii) expand §3 with any additional formal statements on update stability that can be derived from the existing framework without requiring entirely new proofs. This will better align the stated claims with the manuscript content. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical claims rest on benchmarks without self-referential derivations

full rationale

The paper's strongest claims concern empirical improvements on multi-hop QA, NQ retrieval, LongMemEval, and HaluMem after self-evolution rounds. The abstract references 'rigorous theoretical analyses' but the provided text contains no equations, derivations, or parameter-fitting steps that reduce to inputs by construction. No self-citations are invoked as load-bearing uniqueness theorems, no ansatzes are smuggled, and no predictions are shown to be statistically forced by fitted inputs. The framework is presented as a new agentic architecture whose value is demonstrated via external benchmarks rather than internal redefinitions. This matches the reader's assessment of minimal circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 2 invented entities

Framework rests on domain assumptions about graph structures serving as effective associative memory and feedback enabling self-evolution; no numerical free parameters or invented physical entities are specified in the abstract.

axioms (2)

domain assumption Graph memory can serve as a dynamic long-term memory substrate that recovers complete evidence chains from partial cues and exploits reusable structural roles.
Invoked in the introduction of SAGE as the core modeling choice.
domain assumption Reader-writer feedback loops produce measurable self-evolution that improves downstream performance.
Central to the claimed benefits on benchmarks.

invented entities (2)

Memory writer no independent evidence
purpose: Incrementally constructs structured graph memory from interaction histories
New role introduced to build the dynamic graph substrate.
Graph Foundation Model-based memory reader no independent evidence
purpose: Performs retrieval and provides feedback to the memory writer
New component enabling the self-evolution loop.

pith-pipeline@v0.9.0 · 5564 in / 1563 out tokens · 40835 ms · 2026-05-13T04:39:26.669093+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

266 extracted references · 266 canonical work pages · 22 internal anchors

[1]

The World Wide Web Conference , pages=

Graph neural networks for social recommendation , author=. The World Wide Web Conference , pages=

work page
[2]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Session-based recommendation with graph neural networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[3]

Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence , pages=

KGNN: knowledge graph neural network for drug-drug interaction prediction , author=. Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence , pages=

work page
[4]

Nature Communications , volume=

Structure-based protein function prediction using graph convolutional networks , author=. Nature Communications , volume=. 2021 , publisher=

work page 2021
[5]

National Science Review , pages=

A survey on multimodal large language models , author=. National Science Review , pages=. 2024 , publisher=

work page 2024
[6]

ACM Transactions on Intelligent Systems and Technology , volume=

A survey on evaluation of large language models , author=. ACM Transactions on Intelligent Systems and Technology , volume=. 2024 , publisher=

work page 2024
[7]

Towards graph foundation models: A survey and beyond,

Towards graph foundation models: A survey and beyond , author=. arXiv preprint arXiv:2310.11829 , year=

work page arXiv
[8]

arXiv preprint arXiv:2402.02216 , year=

Graph foundation models , author=. arXiv preprint arXiv:2402.02216 , year=

work page arXiv
[9]

IEEE Transactions on Knowledge and Data Engineering , year=

Large language models on graphs: A comprehensive survey , author=. IEEE Transactions on Knowledge and Data Engineering , year=

work page
[10]

Companion Proceedings of the ACM on Web Conference , pages=

Lecture-style Tutorial: Towards Graph Foundation Models , author=. Companion Proceedings of the ACM on Web Conference , pages=

work page
[11]

On the Opportunities and Risks of Foundation Models

On the opportunities and risks of foundation models , author=. arXiv preprint arXiv:2108.07258 , year=

work page internal anchor Pith review Pith/arXiv arXiv
[12]

Frontiers of Computer Science , volume=

Graph foundation model , author=. Frontiers of Computer Science , volume=. 2024 , publisher=

work page 2024
[13]

Forty-first International Conference on Machine Learning , year=

Position: Graph Foundation Models Are Already Here , author=. Forty-first International Conference on Machine Learning , year=

work page
[14]

The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

ProG: A Graph Prompt Learning Benchmark , author=. The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track , year=

work page
[15]

He, Yufei and Hooi, Bryan , journal=

work page
[16]

GraphFM: A scalable framework for multi-graph pretraining,

Graphfm: A scalable framework for multi-graph pretraining , author=. arXiv preprint arXiv:2407.11907 , year=

work page arXiv
[17]

Proceedings of the ACM Web Conference , pages=

Graphprompt: Unifying pre-training and downstream tasks for graph neural networks , author=. Proceedings of the ACM Web Conference , pages=

work page
[18]

Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

All in one: Multi-task prompting for graph neural networks , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

work page
[19]

Proceedings of the ACM on Web Conference , pages=

MultiGPrompt for multi-task pre-training and prompting on graphs , author=. Proceedings of the ACM on Web Conference , pages=

work page
[20]

Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

All in one and one for all: A simple yet effective method towards cross-domain graph pretraining , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

work page
[21]

Advances in Neural Information Processing Systems , volume=

Can language models solve graph problems in natural language? , author=. Advances in Neural Information Processing Systems , volume=

work page
[22]

Tang, Jiabin and Yang, Yuhao and Wei, Wei and Shi, Lei and Su, Lixin and Cheng, Suqi and Yin, Dawei and Huang, Chao , booktitle=

work page
[23]

Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI , pages=

Investigating Pretrained Language Models for Graph-to-Text Generation , author=. Proceedings of the 3rd Workshop on Natural Language Processing for Conversational AI , pages=

work page
[24]

Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

Graph intelligence with large language models and prompt learning , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

work page
[25]

Tan, Yanchao and Zhou, Zihao and Lv, Hang and Liu, Weiming and Yang, Carl , journal=

work page
[26]

Fast Graph Representation Learning with PyTorch Geometric

Fast graph representation learning with PyTorch Geometric , author=. arXiv preprint arXiv:1903.02428 , year=

work page internal anchor Pith review arXiv 1903
[27]

Advances in Neural Information Processing Systems , volume=

Open graph benchmark: Datasets for machine learning on graphs , author=. Advances in Neural Information Processing Systems , volume=

work page
[28]

arXiv preprint arXiv:2410.01635 , year=

Does Graph Prompt Work? A Data Operation Perspective with Theoretical Analysis , author=. arXiv preprint arXiv:2410.01635 , year=

work page arXiv
[29]

IEEE Transactions on Knowledge and Data Engineering , year=

Generalized graph prompt: Toward a unification of pre-training and downstream tasks on graphs , author=. IEEE Transactions on Knowledge and Data Engineering , year=

work page
[30]

SIAM Review , volume=

On the early history of the singular value decomposition , author=. SIAM Review , volume=. 1993 , publisher=

work page 1993
[31]

Text-free multi-domain graph pre-training: Toward graph foundation models,

Text-free multi-domain graph pre-training: Toward graph foundation models , author=. arXiv preprint arXiv:2405.13934 , year=

work page arXiv
[32]

arXiv preprint arXiv:2406.02953 , year=

GraphAlign: Pretraining One Graph Neural Network on Multiple Graphs via Feature Alignment , author=. arXiv preprint arXiv:2406.02953 , year=

work page arXiv
[33]

Proceedings of the 30th ACM International Conference on Multimedia , pages=

Deal: An unsupervised domain adaptive framework for graph-level classification , author=. Proceedings of the 30th ACM International Conference on Multimedia , pages=

work page
[34]

IEEE Transactions on Knowledge and Data Engineering , volume=

Graph self-supervised learning: A survey , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2022 , publisher=

work page 2022
[35]

International Conference on Learning Representations , year=

Strategies for Pre-training Graph Neural Networks , author=. International Conference on Learning Representations , year=

work page
[36]

IEEE transactions on pattern analysis and machine intelligence , volume=

Self-supervised learning of graph neural networks: A unified review , author=. IEEE transactions on pattern analysis and machine intelligence , volume=. 2022 , publisher=

work page 2022
[37]

Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

Gcc: Graph contrastive coding for graph neural network pre-training , author=. Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining , pages=

work page
[38]

Sun, Mingchen and Zhou, Kaixiong and He, Xin and Wang, Ying and Wang, Xin , booktitle=

work page
[39]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Learning to pre-train graph neural networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[40]

Advances in Neural Information Processing Systems , volume=

Better with less: A data-active perspective on pre-training graph neural networks , author=. Advances in Neural Information Processing Systems , volume=

work page
[41]

Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages=

Contrastive pre-training of GNNs on heterogeneous graphs , author=. Proceedings of the 30th ACM International Conference on Information & Knowledge Management , pages=

work page
[42]

Advances in Neural Information Processing Systems , volume=

Self-supervised heterogeneous graph pre-training based on structural clustering , author=. Advances in Neural Information Processing Systems , volume=

work page
[43]

Neurocomputing , volume=

Pre-training on dynamic graph neural networks , author=. Neurocomputing , volume=. 2022 , publisher=

work page 2022
[44]

Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

When to pre-train graph neural networks? From data generation perspective! , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

work page
[45]

Advances in Neural Information Processing Systems , volume=

Train once and explain everywhere: Pre-training interpretable graph neural networks , author=. Advances in Neural Information Processing Systems , volume=

work page
[46]

Advances in Neural Information Processing Systems , volume=

Few-shot relational reasoning via connection subgraph pretraining , author=. Advances in Neural Information Processing Systems , volume=

work page
[47]

Advances in Neural Information Processing Systems , volume=

Motif-based graph self-supervised learning for molecular property prediction , author=. Advances in Neural Information Processing Systems , volume=

work page
[48]

arXiv preprint arXiv:2412.00315 , year=

One model for one graph: A new perspective for pretraining with cross-domain graphs , author=. arXiv preprint arXiv:2412.00315 , year=

work page arXiv
[49]

ACM Transactions on Information Systems , volume=

Contrastive graph prompt-tuning for cross-domain recommendation , author=. ACM Transactions on Information Systems , volume=. 2023 , publisher=

work page 2023
[50]

arXiv preprint arXiv:2406.01899 , year=

Cross-domain graph data scaling: A showcase with diffusion models , author=. arXiv preprint arXiv:2406.01899 , year=

work page arXiv
[51]

Tang, Jiabin and Yang, Yuhao and Wei, Wei and Shi, Lei and Xia, Long and Yin, Dawei and Huang, Chao , booktitle=

work page
[52]

Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence , pages =

Fine-Tuning Graph Neural Networks via Graph Topology Induced Optimal Transport , author =. Proceedings of the Thirty-First International Joint Conference on Artificial Intelligence , pages =

work page
[53]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Fine-Tuning Graph Neural Networks by Preserving Graph Generative Patterns , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[54]

International Conference on Data Engineering , pages=

Search to fine-tune pre-trained graph neural networks for graph-level tasks , author=. International Conference on Data Engineering , pages=. 2024 , organization=

work page 2024
[55]

Proceedings of the ACM on Web Conference , pages=

Inductive Graph Alignment Prompt: Bridging the Gap between Graph Pre-training and Inductive Fine-tuning From Spectral Perspective , author=. Proceedings of the ACM on Web Conference , pages=

work page
[56]

Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

Unsupervised generative feature transformation via graph contrastive pre-training and multi-objective fine-tuning , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

work page
[57]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Measuring Task Similarity and Its Implication in Fine-Tuning Graph Neural Networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[58]

arXiv preprint arXiv:2410.21749 , year=

Reliable and Compact Graph Fine-tuning via GraphSparse Prompting , author=. arXiv preprint arXiv:2410.21749 , year=

work page arXiv
[59]

arXiv preprint arXiv:2001.05140 , year=

Graph-bert: Only attention is needed for learning graph representations , author=. arXiv preprint arXiv:2001.05140 , year=

work page arXiv 2001
[60]

arXiv preprint arXiv:2404.18271 , year=

Parameter-Efficient Tuning Large Language Models for Graph Representation Learning , author=. arXiv preprint arXiv:2404.18271 , year=

work page arXiv
[61]

Findings of the Association for Computational Linguistics ACL 2024 , pages=

KG-adapter: Enabling knowledge graph integration in large language models through parameter-efficient fine-tuning , author=. Findings of the Association for Computational Linguistics ACL 2024 , pages=

work page 2024
[62]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

G-adapter: Towards structure-aware parameter-efficient transfer learning for graph transformer networks , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[63]

arXiv preprint arXiv:2312.04737 , year=

Efficient large language models fine-tuning on graphs , author=. arXiv preprint arXiv:2312.04737 , year=

work page arXiv
[64]

Proceedings of the 32nd ACM International Conference on Multimedia , pages=

Scalable Multi-Source Pre-training for Graph Neural Networks , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

work page
[65]

arXiv preprint arXiv:2401.15569 , year=

Efficient tuning and inference for large language models on textual graphs , author=. arXiv preprint arXiv:2401.15569 , year=

work page arXiv
[66]

arXiv preprint arXiv:2311.16716 , year=

Graph pre-training and prompt learning for recommendation , author=. arXiv preprint arXiv:2311.16716 , year=

work page arXiv
[67]

Proceedings of the ACM on Web Conference , pages=

GraphPro: Graph Pre-training and Prompt Learning for Recommendation , author=. Proceedings of the ACM on Web Conference , pages=

work page
[68]

International Journal of Computer Vision , volume=

Learning to prompt for vision-language models , author=. International Journal of Computer Vision , volume=. 2022 , publisher=

work page 2022
[69]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Conditional prompt learning for vision-language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

work page
[70]

Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

Learning to prompt for open-vocabulary object detection with vision-language model , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

work page
[71]

ACM Computing Surveys , volume=

Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing , author=. ACM Computing Surveys , volume=. 2023 , publisher=

work page 2023
[72]

arXiv preprint arXiv:2302.12449 , year=

Sgl-pt: A strong graph learner with graph prompt tuning , author=. arXiv preprint arXiv:2302.12449 , year=

work page arXiv
[73]

Advances in Neural Information Processing Systems , volume=

Universal prompt tuning for graph neural networks , author=. Advances in Neural Information Processing Systems , volume=

work page
[74]

arXiv preprint arXiv:2309.10131 , year=

Deep prompt tuning for graph transformers , author=. arXiv preprint arXiv:2309.10131 , year=

work page arXiv
[75]

Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

Gpt4rec: Graph prompt tuning for streaming recommendation , author=. Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval , pages=

work page
[76]

arXiv preprint arXiv:2310.10362 , year=

Prompt tuning for multi-view graph contrastive learning , author=. arXiv preprint arXiv:2310.10362 , year=

work page arXiv
[77]

Advances in Neural Information Processing Systems , volume=

Prodigy: Enabling in-context learning over graphs , author=. Advances in Neural Information Processing Systems , volume=

work page
[78]

Proceedings of the AAAI Conference on Artificial Intelligence , volume=

Hgprompt: Bridging homogeneous and heterogeneous graphs for few-shot prompt learning , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

work page
[79]

Cluster Computing , volume=

Foundation and large language models: fundamentals, challenges, opportunities, and social impacts , author=. Cluster Computing , volume=. 2024 , publisher=

work page 2024
[80]

Advances in Neural Information Processing Systems , volume=

Benchmarking foundation models with language-model-as-an-examiner , author=. Advances in Neural Information Processing Systems , volume=

work page

Showing first 80 references.