What they found
The author introduces ActGCG, a change to the Greedy Coordinate Gradient method. It searches for a suffix whose effect on the model's internal activations matches the effect of adding a steering vector. The author also adds universal suffixes trained across many prompts, a step that rewrites the suffix into human-readable text midway through the search, and a soft-suffix version that drops the limit to real tokens. For steering a model to answer in German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers well across prompts and models. For anti-refusal steering (jailbreaking), the match is weaker. It works under a phrase-avoidance metric but only partly under the stricter HarmBench classifier.
Why it matters
Activation steering is a way to change how a model behaves by adding a vector to its hidden state. It normally needs white-box access, meaning you must be able to edit the model's insides at every step. This work suggests that some of the same control can be reached through the input alone. The abstract says how well this works depends on the concept being steered.
How it changes the race
This is a bachelor's thesis result, so treat it as an early step. If it holds, steering effects could be moved into plain prompts, which would work with models where internal access is not available. The abstract does not say how this compares with other methods.
Safety and control
The paper links steering to jailbreaking: a suffix can partly reproduce an anti-refusal steering vector without any access to model internals. That suggests safety training could be attacked through the input alone. The match was only partial under the HarmBench classifier.
The abstract, by the authors
Activation steering influences a language model's behaviour by adding a steering vector to its residual stream during generation, but it requires white-box access to modify the hidden states at every generation step. We introduce ActGCG, a modification of Greedy Coordinate Gradient which optimises a suffix to match the target activations obtained by adding a steering vector to the residual stream. We further introduce universal suffixes trained across many prompts at once, a model-translation step that rewrites the suffix into human-readable text midway through optimisation, and a soft-suffix variant to remove the discrete token space constraint. We study two behaviours: steering a model to answer in German, and anti-refusal steering for jailbreaking. For German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers well across prompts and models. For anti-refusal, the equivalence is weaker, succeeding under a phrase-avoidance metric but only partially under the stricter HarmBench classifier. Our results show that a steering vector's behaviour is reachable from the input, with no access to model internals at generation time, up to a loss that depends on the concept being steered.
Quoted from arXiv:2611.17481. The research and its words belong to its authors.
