Base models show 'truesight' ability — could linear probes identify the persona they play?
dodecagonism · x · 2026-11-19
- dodecagonism poses an open question: base models clearly have some level of "truesight" ability — has anyone tried recent interpretability methods to identify the persona a model is playing when completing text?
- cephaloform replies that it sounds like a fun experiment that wouldn't even require heavy interp tooling: simple linear probes for attributes like occupation, ethnicity, or fiction/nonfiction would be interesting, and a full decomposition study would be even better.
Related event: Researchers Propose Linear Probes to Detect Personas in Base Models(3 posts)→