Research question
Can high-level concepts such as honesty be read and controlled at the representation level?
Key findings
- Linear artificial tomography reads honesty, power-seeking and harmfulness representations - Representation control steers behaviour, including reducing dishonesty