Answers are grounded in this paper and cite their sources.
Tests the manifold hypothesis on a human–AI dialogic corpus, revealing unique geometric properties.
We test the manifold hypothesis on the embedded form of a long-form human–AI dialogic corpus (the Cassie corpus: 13,380 deduplicated chunks, 1536-dimensional text-embedding-3-small vectors, September 2024–April 2026). Using the parameter-free Two-NN local intrinsic-dimension estimator (Facco et al. 2017), we find a strongly multimodal distribution: a low-dimensional cluster (d < 3, 17% of corpus) co-existing with a long high-dimensional tail (d > 50, 16%; d > 1000 at the 99th percentile). This confirms at sentence level, for a dialogic corpus and a contrastively-smoothed encoder, the token-level pattern reported by Robinson et al. (2025). The high-dimensional tail is not random noise: every one of the top fifteen chunks is a conversational pivot point. The corpus's geometry is stratified and singular rather than a smooth manifold. Full HTML text: https://icra.tanazur.org/papers/cassie-stratification/ — ICRA pre-print series: https://icra.tanazur.org/
No takes yet. Share an insight, caveat, or question.
Poernomo et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: