Reproduce “Sweep Less, Steer More: Using Activation Steering Geometry for Reduced Hyperparameter Search” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Nishkal Hundia, Swastik Agrawal, Navita Goyal, Sarah Wiegreffe
The authors use training-activation geometry to rank layers for difference-in-means steering. d′ measures separation of positive and negative examples along the steering direction, requiring no additional forward passes. Across five behaviors and three models, testing the top three layers recovered a median 97% of the best layer’s improvement over baseline in 15 rather than 110 steered runs; validation remains necessary.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.