We introduce an automated activation‑steering approach that plugs into standard labeled datasets—no handcrafted prompt pairs or feature annotation. On 18 tasks and 3 open‑weight models, the introspective variant (iPAS) yields the strongest behavior improvements, and layers on top of ICL/SFT.
We introduce an automated activation‑steering approach that plugs into standard labeled datasets—no handcrafted prompt pairs or feature annotation. On 18 tasks and 3 open‑weight models, the introspective variant (iPAS) yields the strongest behavior improvements, and layers on top of ICL/SFT.
Full write‑up: https://open.substack.com/pub/sashacui/p/painless-activation-steering-pas
Paper: arxiv.org/abs/2509.22739