WRDO
A refusal direction optimization method for safety alignment of large models. It learns aggregation weights for different harmful prompts and forms a composite intervention direction as a weighted sum of per-prompt refusal directions, replacing the equally weighted mean difference vector. It can be used bidirectionally to weaken or restore a model's refusal behavior.
Timeline (last 90 days)
WRDO (Weighted Refusal Direction Optimization) is a method that combines each prompt-specific refusal direction into a composite direction through learned aggregation weights.
The paper compares WRDO against prior single-direction methods on three open-weight language models.
The same optimization mechanism of WRDO can be used not only in the forward direction to weaken refusal, but also in reverse to restore refusal capability in model checkpoints from which refusal behavior has been removed.
Experimental results show that weighted aggregation provides a more effective and more flexible mechanism for manipulating refusal representations, outperforming fixed single-direction interventions.
The paper does not provide universal conclusions across model scales, and the experimental scope of three open-weight models is limited.
All Facts (5)
WRDO (Weighted Refusal Direction Optimization) is a method that combines each prompt-specific refusal direction into a composite direction through learned aggregation weights.
50%UnverifiedThe paper compares WRDO against prior single-direction methods on three open-weight language models.
50%UnverifiedThe same optimization mechanism of WRDO can be used not only in the forward direction to weaken refusal, but also in reverse to restore refusal capability in model checkpoints from which refusal behavior has been removed.
50%UnverifiedExperimental results show that weighted aggregation provides a more effective and more flexible mechanism for manipulating refusal representations, outperforming fixed single-direction interventions.
50%UnverifiedThe paper does not provide universal conclusions across model scales, and the experimental scope of three open-weight models is limited.
50%