IGMiRAG: Intuition-Guided Retrieval-Augmented Generation with Adaptive Mining of In-Depth Memory

ABSTRACT

IGMiRAG introduces a cognitive-inspired retrieval framework that simulates intuition-driven reasoning over hierarchical knowledge graphs. The system constructs a heterogeneous hypergraph to represent multi-granular knowledge structures and relationships. A question parser determines the appropriate depth of retrieval based on query complexity. The model then activates anchor memories and performs bidirectional diffusion through the knowledge graph to retrieve relevant context. Experiments show improved reasoning accuracy and adaptive token efficiency.