Input optimization creates steering prompts that guide language model behavior
Creating Steering Vectors via Input Optimisation
Computation and LanguageMachine Learning
Summary
Language models can be guided by adding special signals inside their calculations, but this usually needs deep access inside the model. The authors developed a way to create input text sequences that cause the model to behave like it was steered, without needing to peek inside during use. They tested this by making the model reply in German or avoid refusing certain prompts. Their method works well for the German case and somewhat for the anti-refusal case, showing that input text alone can mimic steering signals with some limits.
What this means in practice
- •For natural language processing engineers: generate input prompts that emulate internal steering signals to change model behavior without modifying internals at runtime
- •For machine learning platform operators: deploy precomputed input suffixes to steer models in production environments where internal state modification is unavailable
Authors
Batuhan Akar
Abstract
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.