Gespeichert in:
Bibliographische Detailangaben
1. Verfasser: McEntire, Jeremy
Format: Recurso digital
Sprache:
Veröffentlicht: Zenodo 2026
Schlagworte:
Online-Zugang:https://doi.org/10.5281/zenodo.18918719
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Inhaltsangabe:
  • Proves that universal entanglement in transformer representations is a geometric consequence of high-dimensional encoding via the Johnson-Lindenstrauss lemma and concentration of measure. The Entanglement Theorem shows that for any d-dimensional representation with k concept-informative directions, random m-dimensional projections with m > 32k preserve entanglement with probability exceeding 1 - 2exp(-m/128). Experimental validation across GPT-2 to Qwen-7B confirms the theoretical predictions.