Search and Filter

LEKIA: A Framework for Architectural Alignment via Expert Knowledge Injection

Date
Deploying Large Language Models (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.

Research Categories

What is the application?
Who is the user?
Who age?

Publication Source

Publisher
arXiv