Knowledge Graph Adaptive Reasoning with Plan-on-Graph LLM - From Theory to Implementation

Watch a 24-minute technical presentation exploring Plan-on-Graph (PoG), an innovative approach for integrating knowledge graphs with large language models to enhance reasoning capabilities. Learn how PoG addresses key LLM limitations through a self-correcting adaptive planning paradigm that decomposes questions into sub-objectives and enables dynamic knowledge graph exploration. Understand the three core mechanisms - Guidance, Memory, and Reflection - that allow for adaptive navigation, contextual information retention, and error correction in reasoning paths. Examine real-world experimental results demonstrating PoG's superior performance in knowledge graph question-answering tasks, including reduced LLM call frequency and improved accuracy across datasets like CWQ, WebQSP, and GrailQA. Follow along as the presentation covers implementation examples, comparisons with existing methods, and detailed insights into various knowledge graph augmentation techniques, concluding with reflections on data and code applications.