Researchers Map Where "Persona Neurons" Fire in LLMs
A blog post maps where "persona neurons" fire token-by-token across 124 evaluation dialogues, finding chat template tokens carry 25–95% of activations; activating different neurons instantly switches the model's persona.
2026-10-30 ~ 2026-10-31 · 2 related posts
- Persona neurons fire hardest on chat-template tokens, carrying 25–95% of total activation — maxsloef · 2026-10-30
- Steering single 'persona neurons' turns an LLM into Karna, Xuanzang or a lost ghost — Sauers_ · 2026-10-31