Reproduce “Categorical Perception in Large Language Model Hidden States: Structural Warping at Digit-Count Boundaries” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jon-Paul Cacioli

The authors measure whether numeral-format changes create extra separation in large language models’ internal representations. They compare cosine distances between hidden-state centroids with continuous log-magnitude and boundary-boost models. The boundary-boost model fits better at every primary layer in all six tested models, including at 9/10; this is geometric evidence, not proof of human-like perception.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.