Search-on-Graph: Iterative Informed Navigation for Large Language Model Reasoning on Knowledge Graphs
Researchers have developed Search-on-Graph (SoG), a novel method for enhancing large language model reasoning on knowledge graphs. SoG integrates the LLM directly into the path selection process, allowing it to iteratively choose relations based on the reasoning history and available graph structure. This "observe-think-navigate" approach aims to improve accuracy and generalization across different knowledge graph schemas, as demonstrated by its superior performance on six KGQA benchmarks without task-specific fine-tuning. AI
IMPACT Enhances LLM reasoning capabilities on structured data, potentially improving performance in knowledge-intensive AI applications.