MedCoG: Meta-Cognitive Regulation Boosts LLM Medical Reasoning

Researchers propose MedCoG, a Medical Meta-Cognition Agent with Knowledge Graph, to address diminishing returns from inference scaling in large language models (LLMs) for medical reasoning. The system uses meta-cognitive assessments—evaluating task complexity, familiarity, and knowledge density—to dynamically regulate the use of procedural, episodic, and factual knowledge. This on-demand reasoning aims to reduce costs by avoiding indiscriminate scaling and improve accuracy by filtering out distractive knowledge. The approach is validated empirically, with the paper available on arXiv.