The Lattice Representation Hypothesis of Large Language Models

Researchers propose the Lattice Representation Hypothesis, a new framework showing how large language models encode symbolic reasoning through geometric structures. The theory unifies continuous neural representations with formal logic by demonstrating that LLM embeddings naturally form concept lattices that enable symbolic operations through geometric intersections and unions.