phenomenon · first seen 13 Oct 2026
persona vectors
Linear directions in LLM residual-stream activations that encode behavioral traits and allow steering model behavior.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Cross-Lingual Transfer of Evil Vectors Is Directionally AsymmetricPapers · 13 Oct 2026 · Juann Kim, Minku Kang
- Safety Side Effects of Preventative Persona Steering: An Empirical AuditPapers · 13 Nov 2026 · Sabrina Tang, Nikolai Vo, Carson Murillo, Chira Sukumarnpiti, Sarah Wilson
Latest sources
- Safety Side Effects of Preventative Persona Steering: An Empirical AuditPapers · 13 Nov 2026 · Sabrina Tang, Nikolai Vo, Carson Murillo, Chira Sukumarnpiti, Sarah Wilson
- Cross-Lingual Transfer of Evil Vectors Is Directionally AsymmetricPapers · 13 Oct 2026 · Juann Kim, Minku Kang