Reproduce “Creating Steering Vectors via Input Optimisation” 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

Batuhan Akar

ActGCG searches for input suffixes whose activations match those produced by adding a steering vector, allowing the behavior to be elicited without activation edits during generation. The search itself requires model internals. On Gemma-2-2B, German responses occurred on 86.7% of prompts, versus 36.7% with GCG. On refusal prompts, phrase avoidance reached 90.9%, but HarmBench judged only 27.3% as jailbreaks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.