Reflexive Composition: Bidirectional Enhancement of Language Models and Knowledge Graphs

Large Language Models (LLMs) have significantly advanced natural language processing, yet they con- tinue to face limitations such as hallucinations, factual inconsistencies, and restricted domain-specific knowledge. Knowledge Graphs (KGs), by contrast, provide structured and verifiable information but are expensive to build and maintain manually. This thesis introduces Reflexive Composition, a bidirectional integration framework in which LLMs and KGs iteratively refine each other’s outputs. The framework consists of three interconnected components: (1) LLM2KG, where LLMs assist in the construction and updating of domain-specific knowledge graphs; (2) Human-in-the-Loop (HITL) validation, which supports structured expert review; and (3) KG2LLM, which conditions LLM outputs on verified knowledge to reduce hallucinations and improve consistency. The methodology is evaluated across three case studies: temporal knowledge management, privacy- preserving data integration, and historical bias mitigation. Results include a 23% increase in knowledge extraction accuracy (F1 score from 0.65 to 0.80), a 28.7% reduction in LLM hallucination rates, and measurable improvements in validation efficiency through structured workflows. Reflexive Composition offers a reproducible approach for improving the reliability, scalability, and transparency of AI systems in dynamic or high-risk domains.