cs.LGMar 27, 2025

LLaTA: Unlocking Graph Structure Learning with Tree-Guided Large Language Models

Authors: Zhihan Zhang, Xunkai Li, Hongchao Qin, Rong-Hua Li, Guoren Wang

Organizations: Beijing Institute of Technology

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The emergence of large language models (LLMs) has popularized text-attributed graphs (TAGs), creating an urgent need for graph structure learning (GSL) methods that effectively leverage textual information. However, existing GSL approaches are designed for traditional graphs without text, and adapting them to LLMs faces two challenges: defining a suitable optimization objective given LLMs' massive parameters, and designing an efficient architecture without costly fine-tuning. To address these, we propose LLaTA (Large Language and Tree Assistant), which reformulates GSL as a tree optimization framework---shifting from training edge predictors to designing a language-aware tree sampler. LLaTA constructs structural encoding trees via entropy minimization to capture topology, then leverages tree-guided LLM in-context learning to integrate textual semantics without fine-tuning. Extensive experiments on 11 datasets demonstrate LLaTA's flexibility with any backbone, superior scalability over LLM-based GSL methods, and state-of-the-art effectiveness across diverse domains.

Figures & tables

Figure 1: Empirical evaluation on Citeseer with GCN for node classification.
Figure 2: (Left) The overview of LLaTA; (Right) The detailed pipeline of LLaTA, which includes: topology-aware tree prompts , reliable LLM inference and language-aware tree sampler .
Figure 3: Node classification performance on the Cora and Pubmed dataset under real-world scenarios (Sparsity [Edge Removal] and Noise [Edge Addition]).
Figure 4: Hyperparameter analysis of KK , θ\theta and rr . θ\theta and rr are analyzed on the History dataset.
Figure 5: Comparison of training and inference time for LLM-based GSL method.
Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: An illustrative example of hierarchical communities (semantics) in a simple social network.
Figure 7: The overview of our proposed tree-based GSL optimization pipeline and empirical results.
Figure 8: A case study of the tree-prompted LLaTA inference.
Figure 9: Node classification performance on the Citeseer and WikiCS dataset under real-world scenarios (Sparsity [Edge Removal] and Noise [Edge Addition]).
Figure 10: The node classification accuracy of LLaTA with different ϵ\epsilon on Cora, Citeseer, WikiCS and History datasets.
Figure 11: The node classification accuracy of LLaTA with different θ\theta and rr on Cora, Reddit, Child and Photo datasets.
More in this topic

Graph Structure Learning

Graph structure learning infers or adapts network connections from data for statistical discovery and downstream prediction.

See all 134 papers

Explore similar work

CardsList

  1. GaLoRA: Parameter-Efficient Graph-Aware LLMs for Node Classification

    Mar 11, 2026Mayur Choudhary, Saptarshi Sengupta, Katerina PotikaLLMs for Graph LearningText-Attributed Graph Learning

    Department of Computer Science, San Jose State University

  2. Teaching LLMs to See Graphs: Unifying Text and Structural Reasoning

    May 11, 2026Dario VajdaLLM Reasoning with GraphsText-Attributed Graph Learning

    Faculty of Computer and Information Science University of Ljubljana

  3. Position: How can Graphs Help Large Language Models?

    May 4, 2026Xiyuan Wang, Yi Hu, Yanbo Wang +2LLM GroundingKnowledge Augmentation for Language Models

    Institute for Artificial Intelligence, Peking University, Beijing 100871, China · School of Computer Sciences, Beijing University of Posts and Telecommunications, Beijing 100876, China