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

The era of foundation models has revolutionized AI research, yet Graph Foundation Models (GFMs) remain constrained by the scarcity of large-scale graph corpora. Traditional graph data synthesis techniques primarily focus on simplistic structural operations, lacking the capacity to generate semantically rich nodes with meaningful textual attributes-a critical limitation for real-world applications. While large language models (LLMs) demonstrate exceptional text generation capabilities, their direct application to graph synthesis is impeded by context window limitations, hallucination phenomena, and structural consistency challenges. To address these issues, we introduce GraphMaster-the first multi-agent framework specifically designed for graph data synthesis in data-limited environments. GraphMaster orchestrates four specialized LLM agents (Manager, Perception, Enhancement, and Evaluation) that collaboratively optimize the synthesis process through iterative refinement, ensuring both semantic coherence and structural integrity. To rigorously evaluate our approach, we create new data-limited "Sub" variants of six standard graph benchmarks, specifically designed to test synthesis capabilities under realistic constraints. Additionally, we develop a novel interpretability assessment framework that combines human evaluation with a principled Grassmannian manifold-based analysis, providing both qualitative and quantitative measures of semantic coherence. Experimental results demonstrate that GraphMaster significantly outperforms traditional synthesis methods across multiple datasets, establishing a strong foundation for advancing GFMs in data-scarce environments. 2 * Corresponding author 2 Code is available on https://github.com/EnjunDu/GraphMaster. 39th Conference on Neural Information Processing Systems (NeurIPS 2025).

existing connections but cannot create novel nodes or patterns. Node-level mixing techniques like GraphMixup [53] generate synthetic nodes by interpolating features but often produce semantically inconsistent attributes, particularly with textual features. Graph-level synthesis methods such as G-Mixup [18] create entirely new graphs but struggle to balance global structure with local semantic coherence. The core limitation across these traditional methods is their inability to simultaneously preserve meaningful semantics while generating structurally valid expansions-a deficiency particularly pronounced when handling text-attributed graphs (TAGs) where both connectivity patterns and textual node features must remain coherent.

Large language models have demonstrated remarkable capabilities in understanding and generating text [19,11,61,24,2,45], suggesting potential for synthesizing text-attributed graphs. However, directly applying LLMs encounters several critical challenges: standard context windows cannot process entire graphs with numerous textual nodes [3]; LLMs excel at semantic understanding but struggle to maintain structural consistency [12]; and without proper coordination, they tend to produce inconsistent or hallucinated content that fails to capture the intricate balance between topology and semantics [33]. Furthermore, in realistic scenarios with limited available data, LLMs have insufficient examples to learn complex graph patterns [27,44,25].

To address these challenges, we propose GraphMaster, a novel multi-agent framework specifically designed for graph synthesis in data-limited environments. GraphMaster decomposes the complex synthesis task into specialized sub-tasks handled by four collaborative LLM-powered agents, each targeting specific challenges: (1) The Manager Agent coordinates the overall process and determines optimal synthesis strategies based on current graph characteristics, orchestrating the complex synthesis workflow; (2) The Perception Agent analyzes graph structure and employs advanced sampling to identify representative subgraphs processable within LLM context constraints, directly addressing the context window limitations; (3) The Enhancement Agent generates new nodes and edges with consistent semantics and structure, mitigating hallucination by maintaining coherence with existing graph elements; and (4) The Evaluation Agent assesses quality based on both semantic coherence and structural integrity, providing feedback for iterative improvement to ensure structural and semantic consistency. This decomposition enables targeted solutions for each challenge that a single-pass LLM approach cannot address.

Through this collaborative, iterative process, these specialized agents overcome the limitations of both traditional methods and direct LLM applications. The multi-agent architecture enables GraphMaster to effectively balance semantic richness with structural validity-producing high-quality synthetic graph data even with limited training examples. By introducing modular reasoning (through task decomposition), semantic control (via specialized agent expertise),