Why it’s still open
Researchers can now find a direction for ‘being the default Assistant’ and directions for many traits, and models can sometimes tell their own intended outputs from inserted ones. It remains unclear which computations belong to the character rather than to the model as a whole, and early results suggest the persona is mostly active while the model writes, not while it reads.
What would count as an answer
A mechanistic account of which representations are gated or owned by the Assistant persona, tested by interventions: removing or swapping the persona should change exactly the states attributed to it and leave others intact.
Papers that tackle it
- Beckmann Butlin 2026 - Where is the mind — Uses persona-vector research to widen the candidates for an LLM mind to three: the virtual instance, plus two new persona-based views.
- Lu et al 2026 - The Assistant Axis — Maps a space of personas and traits in model activations and finds a leading direction tracking closeness to the default Assistant.
- Chen et al 2025 - Persona vectors — Automatically extracts directions for character traits such as sycophancy and ‘evil’, then uses them to monitor, steer and predict trait shifts caused by fine-tuning.
- Lindsey 2025 - Emergent introspective awareness — Injects known concepts into activations and tests whether models notice and name them.
- Betley et al 2025 - Tell me about yourself — Models fine-tuned on implicit behaviors (risky choices, insecure code) can describe those behaviors without being trained to, and can attribute different learned policies to different personas.
Methods in play
Who’s closest
Placed by the editor, v1 — corrections welcome.
Related questions
- What is the subject
- Whose valence is it
- Writing about a character or becoming it
- How well can LLMs introspect
Know a paper that belongs here, or think this is framed wrongly? Suggest a paper or a correction (forms on the Atlas home).