Internal Geometry

Internal Geometry refers to the implicit, high-dimensional spatial structures and topological mappings that large-language-models (LLMs) develop within their latent spaces to represent semantic relationships, logical constraints, and physical properties. Unlike explicit geometric reasoning, internal geometry emerges as a byproduct of optimizing for predictive accuracy, allowing models to encode complex structural priors without explicit programming.

Core Characteristics

  • Latent Space Topology: Semantic similarity often correlates with geometric proximity in embedding space, forming clusters that reflect conceptual hierarchies.
  • Emergent Structure: Complex reasoning capabilities, such as counting or spatial navigation, arise from the alignment of internal representations with external structural regularities.
  • Implicit Priors: Models internalize physical laws (e.g., conservation of mass, spatial continuity) as geometric constraints within their activation patterns.

Emergent Internal Models

Recent interpretability research highlights how specific internal mechanisms map to geometric or structural understanding:

Implications

  • Interpretability: Identifying these geometric structures allows researchers to probe model behavior and predict failure modes related to spatial or logical reasoning.
  • Generalization: Internal geometry enables zero-shot generalization to novel spatial or structural problems by leveraging learned topological priors.
  • Safety and Alignment: Understanding how models encode physical reality helps in assessing risks related to hallucination in scientific or engineering contexts.

References