INSES: Intelligent Navigation and Similarity Enhanced Search for Knowledge Graph Reasoning

Hang Gao, Dimitris N. Metaxas

Keywords: Knowledge Graph Reasoning, Graph Search, Algorithm, RAG, LLM

TL;DR: We propose INSES, a framework that combines LLM-guided graph navigation with similarity-based expansion and a routing mechanism to enable efficient, accurate multi-hop reasoning over imperfect knowledge graphs.

Abstract: GraphRAG is increasingly adopted for converting unstructured corpora into graph structure, enabling relational, multi-hop reasoning beyond chunk-level retrieval. Most systems then reason over these graphs with classic graph algorithms. However, such traversal, tied to static connectivity and 'connected triple' paths, frequently misses latent semantic links in real-world knowledge graphs (KG) that are noisy, sparse, or incomplete. To address this gap, we introduce INSES (Intelligent Navigation and Similarity Enhanced Search), a dynamic graph-reasoning framework that couples LLM-guided navigation, which prunes noise and steers triple selection with embedding-based similarity expansion to recover hidden links and bridge gaps beyond explicit edges, turning search from a purely structural walk into semantics-aware multi-hop reasoning. Additionally, since GraphRAG style search generally incurs higher complexity than naïve RAG, we complement INSES with a lightweight router that sends simple queries to naïve RAG and escalates complex multi-hop cases to INSES, balancing efficiency and reasoning depth. Across multiple QA benchmarks, INSES consistently outperforms SOTA RAG and GraphRAG baselines. Results highlight complementary strengths of coarse-grained text retrieval for easy cases and fine-grained triple reasoning for harder ones. On the MINE benckmark, INSES remains robust across KGs produced by KGGEN, GraphRAG, and OpenIE, improving accuracy by 5\%, 10\%, and 27\%. This work opens the door to adaptive, router-backed KG reasoning.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 3812