AutoLexSteer: Automatic Contrast Construction for Lexical Activation Steering

Researchers introduced AutoLexSteer, described as the first fully automated process for building steering vectors for large language models, alongside a new steering vector type called Hangman that operates using word senses. AutoLexSteer uses families of closely related words extracted from WordNet to specify a steering source to avoid and a desired steering target, enabling steering at the level of individual words and sets of word senses and the ability to steer LLM behaviors such as sycophancy. The dataset and code are available at https://github.com/ShuheWang1998/autolexsteer, per the arXiv paper 2609.06879v1.

arXiv:2609.06879v1 Announce Type: cross Abstract: Steering vectors have rapidly emerged as a popular and effective method for guiding the output of LLMs in very specific ways. But constructing accurate steering vectors is a difficult manual process due to the opacity of embeddings. We introduce Hangman, a novel type of steering vector that operates using word senses, as well as AutoLexSteer, the first fully automated process for building steering vectors. AutoLexSteer employs families of closely-related words extracted from WordNet to specify both the steering source to be avoided and the desired steering target. The steering vectors are quite precise, can be used to steer at the level of words and sets of word senses meanings , and are able to steer certain LLM behaviors like sycophancy. The dataset and code can be found at https://github.com/ShuheWang1998/autolexsteer.