In Situ Graph Reasoning and Knowledge Expansion Using Graph‐PRefLexOR

The pursuit of automated scientific discovery has evolved from symbolic logic to modern AI, advancing reasoning and pattern recognition. While large language models (LLMs) excel in fluency and recall, they often struggle with structured reasoning and symbolic abstraction. We introduce Graph‐PRefLexOR (Graph‐based Preference‐based Recursive Language Modeling for Exploratory Optimization of Reasoning), a framework that extends transformer‐based LLMs with explicit graph construction, symbolic abstraction, and recursive refinement during answer generation. This architecture integrates graph‐based representations into language generation, enabling interpretable, multistep reasoning. Inspired by reinforcement learning and category theory, Graph‐PRefLexOR defines reasoning as a structured mapping from a user task to knowledge graph , patterns , and answer . Concepts become nodes, and relationships form edges, supporting hierarchical inference and adaptive learning. Demonstrations span hypothesis generation, materials design, and creative reasoning—such as linking mythological ideas like “thin places” to materials science—showcasing generalization beyond training domains. We propose a “knowledge garden growth” strategy to foster interdisciplinary insights. A 3‐billion‐parameter Graph‐PRefLexOR outperforms standard LLMs in reasoning depth, structure, and applicability. This work lays the foundation for general‐purpose, interpretable reasoning systems that can drive scientific discovery.