GraphVulcan: Discretizing Graph Structures into Tokens to Enable Graph Reasoning for Large Language Models

Key Information About the GraphVulcan Framework

• Development Team: Alibaba Behavioral Risk Control Team
• Core Technology: Discrete graph tokenization technology, which converts graph structures into token sequences
• Problem Solved: Breaking through the limitation of large language models (LLMs) in understanding non-Euclidean graph-structured data
• Academic Achievement: The related research was accepted by SIGKDD 2026, a top data mining conference
• Open-Source Address: GitHub Repository

This framework aims to enable LLMs to understand graph structures like they process text through tokenization, thereby achieving structural reasoning.

Graph-structured data is widely present in fields such as social networks, molecular structures, and recommendation systems. However, large language models are inherently good at sequential text and have limited understanding of non-Euclidean structures.

Traditional methods like text descriptions or adjacency matrices either lose structural information or are difficult for LLMs to understand effectively. The Alibaba team proposed the GraphVulcan framework to address this pain point.

The core of GraphVulcan is discrete graph tokenization, which has three key advantages:

1. Graph Structure Encoding: Design a graph vocabulary (graph_vocab) to map nodes, edges, and relationships into discrete token sequences while preserving topological structures.
2. Next Graph Token Prediction: Draw on the next-token prediction paradigm of LLMs to learn structural patterns by predicting the next token in the graph sequence.
3. Structure-Aware Reasoning: Enable LLMs to understand complex relationships between nodes and edges as well as multi-hop connections through training.

The GraphVulcan open-source codebase includes the following modules:

• ds_config/: Training and dataset configuration
• evaluate/: Evaluation scripts and metric calculation
• gen_data/: Data generation and processing tools
• graph_vocab/: Graph vocabulary construction and management
• scripts/: Training and inference scripts
• utils/: Utility functions

The modular design facilitates the reproduction of paper results and extended research.

This framework has important applications in multiple fields:

1. Risk Control and Anti-Fraud: Identify abnormal patterns and detect fraud rings (a core business scenario of Alibaba).
2. Knowledge Graph Reasoning: Enhance the performance of LLMs in graph completion and multi-hop reasoning.
3. Molecular Materials Science: Accelerate scientific computing such as drug discovery and material design.
4. Recommendation Systems: Improve the understanding of user-item interaction graphs and optimize recommendation quality.

The research results related to GraphVulcan were accepted by SIGKDD 2026, with the paper title "Towards Next Graph Token Prediction: Discrete Graph Tokenization for Structural Reasoning in Large Language Models".

This work not only provides theoretical innovation but also offers a reproducible research foundation for academia and industry through open-source code and data pipelines.

GraphVulcan transforms graph reasoning into a token prediction task, opening up a new direction for the integration of graph neural networks and LLMs.

• For developers: Provides a complete toolchain to quickly get started with graph structure language modeling.
• For researchers: Offers an extensible framework to explore complex graph reasoning tasks.

In the future, the deep integration of graph structures (as an important modality) with LLMs will become a trend, and GraphVulcan has laid the technical foundation.