An adaptive algorithm for finding the right steering coefficient
Interpretability-based safety audits using universal steering and representation engineering, with an adaptive method for locating optimal activation-steering coefficients. Robustness varied sharply across models.
Research on interpretability-based safety audits applies universal steering and representation engineering, introducing an adaptive algorithm to identify optimal activation-steering coefficients for unsafe behavioural concepts. Evaluation found contrasting robustness across models — some highly vulnerable to jailbreaking, others robust.
The adaptive part is the contribution. Activation steering works by adding a direction to the model's internal state, and the coefficient — how hard to push — has been set by hand, per model, per concept. Too little does nothing; too much degrades the model into incoherence. Automating that search is what turns a research technique into an audit tool.
Used offensively, that is a stronger attack than prompt-based jailbreaking: it requires weight access, but it operates on the mechanism directly rather than searching for phrasings that happen to work. Used defensively, it is a way to measure how far a model can be pushed before its refusals break.
The contrasting-robustness finding is the practically useful result and the one that needs the most care. Models differing sharply in steerability suggests something about training — safety training that shapes internal representations rather than only output behaviour would plausibly resist steering better. But the paper compares a small set, and "some models were robust" invites over-reading.
What makes this the right direction is that it tests the mechanism rather than the surface — the same complaint that behavioural evaluation only ever returns a lower bound. A steering audit at least establishes what the representation permits.
arXiv — Breaking Bad: Interpretability-Based Safety Audits of State-of-the-Art LLMs → · arXiv — Towards Automated Circuit Discovery for Mechanistic Interpretability →