An ELM-Inspired Dual-Layer Control Architecture: Regulating Predictive Errors in Large Language Models and AI Hallucinations

This article outlines a theoretical approach to reducing predictive errors and hallucinations in Large Language Models (LLMs) by utilizing a dual-layer overarching control architecture inspired by the Empathic Logic Model (ELM). Before data retrieval, the system conducts a pre-processing prompt inspection to ensure the user's Operational What and Contextual Why are present and proportionally aligned. After Retrieval-Augmented Generation (RAG) occurs, the AI performs a subsequent proportional alignment check to determine if the retrieved sources match with each other. If either alignment check fails, the AI securely halts the generative process and utilizes ELM modes to clarify the missing context or resolve source discrepancies directly with the user. By gaining this clear understanding and additional data, the AI can safely and accurately redo the process.