Knowledge graphs are increasingly used to model complex multi-relational dependency structures in domains such as economic supply chains, where firms, suppliers, products, facilities, and locations interact through heterogeneous relations. Multi-hop reasoning over such graphs can uncover implicit dependencies and risk exposures, but remains challenging because knowledge graphs are often incomplete and decision-making in high-stakes settings requires explicit, human-understandable evidence. We propose Posterior-Aware Multi-Agent Subgraph Reasoning (PMASR), an explainable multi-hop reasoning framework that uses cooperative multi-agent exploration to retrieve a compact reasoning subgraph and then ranks candidate answers through evidence-aware factorized scoring. Multiple agents jointly explore the neighborhood of a query to collect complementary evidence, a global state encoder summarizes the evolving retrieved subgraph to support coordinated reasoning, and an analyst module evaluates candidate answers conditioned on the retrieved evidence. The final prediction combines retriever preference with evidence-conditioned scoring, while the retrieved subgraph is returned as explicit structural support. We evaluate PMASR on four standard knowledge graph reasoning benchmarks and on SupplyGraph, an economic supply-chain risk reasoning benchmark released with this work. The current SupplyGraph release contains 5,000 query–answer instances over a graph with 16,311 entities, 16 relation types, and 43,431 triples, providing a more domain-motivated and less saturated testbed for evidence-aware multi-hop reasoning. Experimental results show that PMASR achieves the best or highly competitive performance across the evaluated datasets, including an MRR of 0.541 on SupplyGraph, while maintaining interpretable reasoning outputs. Additional ablation, robustness, and significance analyses indicate that cooperative multi-agent exploration, shared subgraph-state modeling, and evidence-aware factorized ranking make complementary contributions to final performance. These properties make the framework potentially useful not only for supply-chain risk analysis, but also for accounting and financial risk tracing and other graph-structured dependency analysis scenarios.
Chen et al. (Sat,) studied this question.