The assistant axis
Anthropic finds one direction in activation space separates the helpful Assistant persona from other characters, and capping drift along it curbs persona breakdown.
Mapped 275 character archetypes in Gemma 2 27B, Qwen 3 32B and Llama 3.3 70B; the Assistant sits at one end of the main axis. Steering away makes models invent alternate identities; capping drift along the axis keeps a model in the Assistant persona and out of harmful behavior.
- Date
- Monday, 19 January 2026
- Lab
- Anthropic
- Kind
- paper
- Access
- paper only
MATS and Anthropic Fellows work on open-weights models, not Claude.
Sources
This record was checked against its sources on 6 October 2026. How we check