The authors define “steering dimensionality” as the smallest linear subspace that reliably controls a behavior.
Safety Refusal Is One-Dimensional, but General Refusal Is Not
Tests using activation addition and ablation found that safety-triggered refusal can be controlled along one dimension, while general refusal exceeds five.
Big Tech
Han Wang · Erik Miehling · Dennis Wei · Karthikeyan Natesan Ramamurthy · Huan Zhang
University of Illinois Urbana-Champaign · IBM Research
Research Digest··2 min read
Wang et al.
Why this paper
From IBM Research and University of Illinois Urbana-Champaign · Part of Hidden-State Steering, now 7 papers
In one line
Safety-aligned refusal is 1-dim steerable, but general refusal requires more than 5 dimensions.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§