park-2025-iclr-categorical-polytope-geometry
IN premise — summaries/2026/08/24/park-2024-categorical-hierarchical-concepts-s2-using-this-result-we-show-that-semantic-hierarchy-between-co-chunk-1.md
Created 2026-08-25T02:58:24+00:00
Park et al. (ICLR 2025) prove that categorical concepts in LLM representation spaces are geometrically represented as polytopes (convex hulls of vertex vectors), with 'natural' concepts forming (k−1)-simplices.
Summary
This is a proven result showing that the way large language models organize categorical concepts (like "dog," "bird," "car") in their internal math space isn't arbitrary — the concepts sit at the corners of a geometric shape (a polytope), and the categories humans find most "basic" form the simplest possible version of that shape. It matters because it gives a concrete, checkable geometry for what the model "knows" about a category, turning fuzzy questions about representation into something you can measure, test, and potentially manipulate.
Dependents
These beliefs depend on this one:
- OUT sae-functional-abstraction-extends-geometric-scope — SAE features activating on functional analogies (transit feature on wormholes) and cross-modal inputs (text-trained features firing on images) demonstrate the geometric space encodes intensional and relational structure beyond Park's extensional categorical polytopes, broadening the geometric framework's explanatory scope to include non-lexical, compositional semantics.
- OUT sae-neighborhood-as-polytope-navigation — SAE feature neighborhoods (e.g., Golden Gate Bridge → Alcatraz → San Francisco → California) are the operational navigation algorithm for the categorical polytope geometry: each SAE feature is a polytope vertex, the neighborhood structure is the polytope edge adjacency, and cross-model universality confirms this polytope is a shared semantic object rather than a model-specific artifact.