The paper introduces Multi-Agent Agentic Graph Learning (MAAGL), a framework designed to improve performance on graph reasoning tasks. Existing agentic graph learning methods often rely on a single agent or a shared reasoning policy across the entire graph, which can be suboptimal when graphs exhibit heterogeneous patterns. MAAGL addresses this by partitioning the graph into communities and assigning an independent agent to each community, enabling region-specific specialization. The framework represents structural and semantic evidence separately. Structural evidence is summarized using dynamically updated signatures that are permutation-invariant and fixed in size. Semantic evidence is filtered to the top-k nodes ranked by relevance. MAAGL utilizes historical trajectories with similar signatures to estimate agent confidence and triggers debate-style collaboration when needed. Experiments on four benchmark datasets demonstrate that MAAGL outperforms state-of-the-art AGL methods.
Source: https://arxiv.org/abs/2609.09565