LEKIA: A Framework for Architectural Alignment via Expert Knowledge Injection (2507.14944v1)

Abstract: Deploying LLMs in high-stakes domains is impeded by a dual challenge: the need for deep, dynamic expert knowledge injection and nuanced value alignment. Prevailing paradigms often address these challenges separately, creating a persistent tension between knowledge and alignment; knowledge-focused methods like Retrieval-Augmented Generation (RAG) have limited deep alignment capabilities, while alignment-focused methods like Reinforcement Learning from Human Feedback (RLHF) struggle with the agile injection of expert wisdom. This paper introduces a new collaborative philosophy, Expert-owned AI behavior design, realized through Architectural Alignment-a paradigm that unifies these two goals within a single framework called the Layered Expert Knowledge Injection Architecture (LEKIA). LEKIA operates as an intelligent intermediary that guides an LLM's reasoning process without altering its weights, utilizing a three-tiered structure: a Theoretical Layer for core principles, a Practical Layer for exemplary cases, and an Evaluative Layer for real-time, value-aligned self-correction. We demonstrate the efficacy of this paradigm through the successful implementation of a LEKIA-based psychological support assistant for the special education field. Our work presents a path toward more responsible and expert-driven AI, empowering domain specialists to directly architect AI behavior and resolve the tension between knowledge and alignment.