Reproduce “Token embeddings violate the manifold hypothesis” 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
Michael Robinson, Sourya Dey, Tony Chiang
This research develops and applies novel statistical tests, grounded in differential topology, to empirically demonstrate that Large Language Model token embeddings violate the manifold hypothesis. It reveals that these embeddings exhibit irregular topological structures with non-constant local dimensions, challenging foundational assumptions about their organization.
Key insights
- LLM token embeddings consistently violate the manifold hypothesis due to highly variable local intrinsic dimensions across their embedding spaces.
- Many token neighborhoods also reject the more generalized fiber bundle hypothesis, exhibiting increasing local dimension estimates as the neighborhood radius expands, which contradicts theoretical expectations for fiber bundles.
- The distinct patterns of topological irregularities observed in different LLMs, even those with identical token sets, suggest that training methodologies imprint unique "topological fingerprints" on their embedding spaces.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.