The Lattice Representation Hypothesis of Large Language Models

In three linesA hypothesis proposes that LLMs encode concept lattices in their embedding geometry. The framework unifies the Linear Representation Hypothesis with Formal Concept Analysis (FCA), showing that linear attribute directions induce lattices via half-space intersections. Experiments on WordNet validate that embeddings capture logical and hierarchical structures.

Summary generated by Claude — human-verified