Skip to content
InfoResearchPreprintLLM-specific

Creating Steering Vectors via Input Optimisation

Published
Record updated
View JSON

Summary

Researchers show that the behaviour of an activation steering vector can be reproduced through input alone. Their method, ActGCG, modifies Greedy Coordinate Gradient to optimise a suffix whose activations match those produced by adding a steering vector to the residual stream. For steering a model to answer in German, a discrete suffix reproduces the steering vector's behaviour almost perfectly and transfers across prompts and models. For anti-refusal steering, the match is weaker, succeeding under a phrase-avoidance metric but only partially under the HarmBench classifier.