Research map / Multi-agent systems and games / LLM agents

GraphMaster: Automated Graph Synthesis via LLM Agents in Data-Limited Environments

NeurIPS 2025 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GraphMaster, a four-agent LLM framework (Manager, Perception, Enhancement, Evaluation) for synthesizing text-attributed graph nodes in data-limited settings, evaluated on new "Sub" dataset variants with GNN downstream tasks.

Problem

Graph foundation models are limited by scarce large-scale text-attributed graph data, and existing graph synthesis methods fail to generate semantically coherent textual node attributes while preserving structure; direct LLM generation faces context limits, hallucination, and structural inconsistenc

Approach

A hierarchical RAG-style multi-agent loop where a Manager picks a semantic or topological synthesis mode, a Perception agent samples a context subgraph (community detection plus mode-conditioned PPR) that fits the LLM context, an Enhancement agent generates nodes and edges, and an Evaluation agent filters them with an adaptive threshold until convergence.

Key results

• On GCN, GraphMaster reaches 93.7% accuracy on Cora and 88.3% on Citeseer, versus 91.2% for the best baseline (LLM4RGNN) on Cora and 84.5% (Synthesis-LLM) on Citeseer.
• Human traceability score of 0.92 for GraphMaster versus 0.66 for Mixed-LLM and 0.59 for Synthesis-LLM.
• Ablation: removing the Perception Agent drops average performance by 5.2% and removing the Evaluation Agent by 4.1%.

Measured effects reported in the paper

Surprising findings

• Several classic graph synthesis baselines fall below the original-data baseline when using Sentence-BERT features.
• DeepSeek-R1-32B ranks second among LLM backbones despite LLaMA-33B having more parameters, which the authors read as architecture and pretraining mattering more than size.

Limitations

• Evaluation uses only GNN downstream accuracy and F1 on the authors' own derived Sub datasets; no external benchmark for synthesis quality.
• Human evaluation and Grassmannian coherence scores are reported without a detailed protocol in the main text.
• Theoretical guarantees are claimed in the appendix comparisons but not substantiated in the provided text.

Experimental setup

Datasets: Cora Citeseer Wikics History Arxiv2023 Children SubCora SubCiteseer SubWikics SubHistory SubArxiv2023 SubChildren

Benchmarks: Cora Citeseer Wikics History Arxiv2023 Children

Metrics: accuracy F1 score degree distribution KS test clustering coefficient similarity traceability score

Baselines: GAugO GraphEdit LLM4RGNN GraphSmote G-Mixup IntraMix GraphAdasyn FG-SMOTE AGMixup GAG LLM4NG Mixed-LLM

Builds on: retrieval-augmented generation GAG LLM4NG GraphSMOTE G-Mixup

Closest papers by meaning

• GraphArena: Evaluating and Exploring Large Language Models on Graph Computation (ICLR 2025 · similarity 0.89)
• Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.88)
• Can Large Language Models Analyze Graphs like Professionals? A Benchmark, Datasets and Models (NeurIPS 2024 · similarity 0.88)
• How Do Large Language Models Understand Graph Patterns? A Benchmark for Graph Pattern Comprehension (ICLR 2025 · similarity 0.88)
• Graph-of-Agents: A Graph-based Framework for Multi-Agent LLM Collaboration (ICLR 2026 · similarity 0.88)
• Attention Mechanisms Perspective: Exploring LLM Processing of Graph-Structured Data (ICML 2025 · similarity 0.88)