WRDO learns weighted directions to change LLM refusal behavior
Original title: WRDO: Weighted Refusal Direction Optimization in Aligned Language Models
The method is tested on three open-weight models with HarmBench and can also be reversed to restore refusal.
The paper introduces Weighted Refusal Direction Optimization, which builds a composite refusal direction from per-prompt directions using learned weights instead of one difference-of-means vector. It compares WRDO with prior single-direction methods on three unnamed open-weight language models, using three guard judges and HarmBench, and checks utility on standard benchmarks. The same optimization is also run in reverse to restore refusal in an abliterated checkpoint. The authors report more effective manipulation of refusal representations.
Unconfirmed: Effectiveness claims are the authors' reported results; the three models are not named.
- Harmful prompts can receive unequal weights in the final direction.
- Evaluation uses three guard judges plus HarmBench.
- Utility is measured with standard capability benchmarks.
- A reverse setting aims to restore refusal after abliteration.
From the original article
1 min read at arxiv.orgAligned large language models are trained to comply with benign requests while refusing harmful ones. However, recent work shows that this refusal behavior is fragile and can often be disrupted by simple activation- or weight-space interventions.
By Grigoriy Evseev · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.