Reproduce “WRDO: Weighted Refusal Direction Optimization in Aligned Language Models” 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
Grigoriy Evseev, Vladimir Solodkin, Dmitry Bylinkin, Daniil Medyakov, Vladimir Aletov, Aleksandr Beznosikov
WRDO learns to combine prompt-specific directions for editing language-model refusal behavior, rather than averaging them uniformly. It optimizes weights for directions derived from harmful prompts, then uses their combination in a rank-one weight intervention. On Qwen3-8B-Base at α=3.0, GRPO-IS reached a WildGuard-measured attack success rate of 0.93, versus 0.002 for uniform aggregation; results cover three models and rely on automated judges.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.