MAAGL Partitions Graphs into Communities with Fixed-Size Structural Signatures

The paper introduces MAAGL to fix the shared-policy bottleneck in existing agentic graph learning. Graphs are partitioned into communities; each receives an independent LLM agent holding its own memory. Structural evidence is compressed into a dynamically updated signature that remains invariant to node order and bounded in length, while semantic context is truncated to top-k relevant nodes.

Experiments report consistent gains over single-agent and role-based AGL baselines across four standard datasets. Ablations confirm that signature-based confidence estimation reduces context length growth and that debate-style collaboration activates only when signature similarity to prior trajectories falls below an internal threshold.

Prior AGL work verbalized entire neighborhoods, breaking permutation invariance and inflating token counts. MAAGL decouples structure from semantics and localizes policy learning, aligning with observed community structure in citation, social, and molecular graphs.

Operational deployment requires mapping real-world graphs to the same community detection step used in the benchmarks and maintaining signature update rules across streaming edge arrivals.