Decomposing Complexity: A Difficulty-Aware Multi-agent Framework for Open-Domain Knowledge Graph Construction

Knowledge graphs (KGs) connect information and reveal hidden relationships in various domains. Recent multi-agent research on KG construction offers a more scalable alternative to traditional methods. However, the community widely underestimates the difficulty of two foundational tasks: open information extraction (IE) and automated schema induction. Our quantitative analysis reveals that these tasks are far more difficult than commonly perceived, causing even frontier larger language models (LLMs) to lack consistent reliability. This finding indicates that an end-to-end approach relying on simple, direct prompting for these tasks is insufficient to build robust systems. In this paper, we propose a difficulty-aware multi-agent framework (DAMA), which tackles the difficulty of KG construction through strategic decomposition. DAMA deconstructs these key tasks into more tractable sub-tasks, allowing a team of diverse agents to solve them collaboratively. This decomposition tailors each agent’s task space, which reduces the required context length and mitigates semantic reasoning demands. For a more comprehensive evaluation of factual correctness, we also introduce two novel metrics: contradiction rate (CR) and substantive information density (SID). Experiments on the MINE benchmark and open-domain datasets demonstrate that DAMA outperforms comparative methods, enabling the efficient construction of high-quality KGs from open-domain documents.