This research reveals geometric structures of persona vectors in language models, suggesting a foundational understanding of their traits.
Anthropic's persona vectors (2025) — activation patterns controlling traits like sycophancy, hallucination, and evil in language models — are shown to be projections of the Berry monopole structure (HP40, c₁=1) on the Eckert manifold. The three core persona dimensions decompose into opacity (O), reactivity (R), and coupling (α). The Berry connection A = 1−2O predicts the nodal line between safe and unsafe persona regions. Steering with persona vectors is geometric parallel transport on the Hopf bundle. The framework provides the missing theoretical foundation for why persona vectors exist, what determines their geometry, and why certain traits cluster.
No takes yet. Share an insight, caveat, or question.
Anthony W. Eckert (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: