WRDO: Weighted Refusal Direction Optimization for Large Language Model Alignment
Start here
The paper proposes WRDO (Weighted Refusal Direction Optimization), which uses learned aggregation weights to combine multiple refusal directions computed per prompt into a composite direction for refusal intervention in large language models, avoiding the difference-of-means direction averaging heterogeneous harmful prompts into a single direction and conflating the refusal phenomenon. The authors compare it with existing single-direction methods on three open-weight models using three guardrail judges and HarmBench, and measure capability impact with general capability benchmarks.
Generated by AI from reports · Updated 22 days ago
Attention and developments
Attention, hourly
Now 31, 0% in 24 hours·Peak 31 (Nov 6 03:00)
Related research
This event's paper
- WRDO: Weighted Refusal Direction Optimization in Aligned Language Models
arXiv · 2026-10-21 · Deep reading
Progress timeline
- arXiv: Jailbreaks, Prompt Injection, Poisoning & DefensesWRDO: Weighted Refusal Direction Optimization in Aligned Large Language Models
The researchers propose WRDO (Weighted Refusal Direction Optimization), which uses learned aggregation weights to combine multiple per-prompt refusal directions into a composite direction for refusal intervention in aligned large language models. The paper points out that the commonly used mean-difference direction averages heterogeneous harmful prompts into a single direction, potentially conflating distinct refusal-related phenomena, whereas WRDO allows different harmful prompts to contribute unequally to the final intervention direction. The authors compare against existing single-direction methods on three open-weight models, using three guardrail judges and HarmBench as external evaluations, and measure the intervention's impact on model capability with general-capability benchmarks. The paper also examines a defensive use, applying the same optimization mechanism in reverse to restore refusal behavior in model checkpoints whose refusal direction has been removed (abliterated); the authors' results show that weighted aggregation is a more effective and flexible mechanism for manipulating refusal representations.