High-Dimensional Random Projection for Activation Steering in Language Models

In three linesHiDRA, a training-free activation steering method, uses high-dimensional random projection to improve behavioral control of LLMs. It outperforms linear difference-in-means approaches by capturing discriminative signals in nonlinear feature subspaces, with consistent gains across multiple model families.

Summary generated by Claude — human-verified