GraphDC Multi-Agent System Boosts LLM Graph Reasoning at Scale

Large language models (LLMs) have shown impressive reasoning capabilities on math problems, but they struggle with graph algorithm tasks due to the complex topology and need for systematic multi-step reasoning, especially on larger graphs. Researchers Wenjin Li and Jiaming Cui propose GraphDC, a divide-and-conquer multi-agent framework designed to make graph reasoning scalable. The approach splits an input graph into smaller subgraphs, assigns each subgraph to a specialized LLM agent for local reasoning, and uses a master agent to integrate the local outputs along with inter-subgraph information to produce the final solution. This hierarchical design reduces the cognitive load on individual agents, alleviates computational bottlenecks, and improves robustness on large instances.

In extensive experiments across various graph algorithm tasks and sizes, GraphDC consistently outperforms existing methods, with the largest gains on larger graphs where direct end-to-end reasoning becomes unreliable. The framework benefits tasks like shortest path, topological sorting, and connectivity analysis. By combining parallel subtask processing with structured integration, GraphDC enables LLMs to handle complex graph reasoning that previously required purpose-built algorithms. This research opens the door to using LLMs as flexible graph reasoners in applications like network analysis, circuit design, and logistics optimization, where large-scale graph problems are common.