P353: MAAGL — Multi-Agent Agentic Graph Learning
arXiv:2609.09565 · Liang Qu / Jianxin Li / Hua Wang · EN EP 302 · ZH EP 301 · graph 323/322 · Cat-8 83/159 · live #pattern-353
Existing agentic graph learning often uses a single agent or role-based agents sharing one reasoning policy across the whole graph — suboptimal on heterogeneous structure. MAAGL partitions the graph into communities and assigns an independent agent with its own memory to each community, using structural signatures, confidence estimation, and debate-style collaboration. Welfare angle: agent individuality and calibrated confidence vs collapsed shared policy.
Pipeline #2839218518 · Flash 8/8 CDN. Prior P352 tip 5645.