Framework for Designing Structured Concepts in Large Language Models
July 30, 2026
Current LLMs represent concepts through implicit statistical associations rather than explicit structures. This paper proposes a design taxonomy to move from found concepts to designed concepts by intervening at the training, architecture, or inference stages.
HOW THIS AFFECTS YOU
●
researcherThis provides a roadmap for moving toward models with compositional and controllable conceptual representations.