ActGCG mimics activation steering from the prompt alone
A new arXiv paper describes a method that optimises input text to reproduce what activation steering does, without touching model internals during generation.
In plain words
Activation steering changes how a language model behaves by editing its internal signals while it writes. The paper's method, ActGCG, instead tunes a piece of input text so the model behaves as if that internal edit had happened. This worked well for switching output language, but less completely for bypassing refusals.
What was announced
The authors present ActGCG, a variant of the Greedy Coordinate Gradient attack that searches for a text suffix whose effect on the model matches the one produced by adding a steering vector to the residual stream. They also train suffixes that work across many prompts at once, rewrite the suffix into readable text partway through optimisation, and offer a soft-suffix version that relaxes the discrete token constraint.
Two behaviours were tested. A discrete suffix made a model answer in German almost as reliably as the steering vector itself, and it carried over to other prompts and models. Anti-refusal steering, used for jailbreaking, transferred less cleanly: it passed a phrase-avoidance check but only partly satisfied the stricter HarmBench classifier. The authors conclude that steering behaviour can be reached from the input alone, with an error that varies by concept.
What it does not show
- The evidence is the paper's abstract only; full experimental details, model sizes and datasets are not visible in this excerpt.
- Anti-refusal transfer was only partial under the HarmBench classifier, so equivalence to steering is not uniform across behaviours.
- No independent party has reproduced these results; findings rest on the authors' own evaluation.
For security and IT teams
- Scope
- The method requires no access to model internals at generation time, which the authors say makes steering-like behaviour reachable from input text alone.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.