MAAGL Uses Specialized Agents for Structure-Aware Graph Learning

Summary

The paper introduces MAAGL, a multi-agent agentic graph learning framework for graph reasoning with large language models. Existing approaches commonly apply one shared reasoning policy across graph regions, which may be poorly suited to graphs with heterogeneous structures and semantics. MAAGL partitions a graph into communities and assigns an independent agent to each community, enabling region-specific specialization. It represents structural and semantic evidence separately: a dynamically updated, fixed-size structural signature preserves permutation invariance, while semantic evidence is reduced to the top-k most relevant nodes. Agents use historical trajectories with similar signatures to estimate confidence and initiate debate-style collaboration when confidence is insufficient. Experiments on four benchmark datasets show that MAAGL outperforms state-of-the-art agentic graph learning methods.