Activation steering is a technique for controlling large language models at inference time without retraining. A new preprint on arXiv argues that the method's effectiveness hinges on the choice of reference data used to compute the steering direction.
The authors propose 'genesis steering,' which derives a genesis vector from contrastive examples in the training data. The abstract does not specify which datasets or models were used, nor does it report quantitative results; it stops at the introduction of the method.
Because the source is a single truncated abstract, there are no independent results to compare. The claim is limited to the proposal itself: that grounding the steering vector in training-data contrastive examples can improve activation steering.