Preprint demonstrates correction for unintended effects of language model steering

Researchers report in a preprint that a second steering vector corrects about 78% of unintended changes in generated text while preserving the intended behavior in about 93% of generations. The intervention targets side effects of activation steering, a technique that modifies a language model's internal activations during generation without retraining its weights or editing its prompt. The study finds that steering systematically changes features such as formality, use of evidence and ambiguity. Its complementary vector targets the affected feature; the correction preserves the intended behavior in about 93% of generations.