Revisiting Over-refusal in Safety-aligned Large Language Models: A Representation and Optimization Perspective
Traces over-refusal to imbalanced gradient coupling, then orthogonalizes gradients with contrastive learning.
Explainer · on the wire
The explainer for this story is still on the wire.
The claim, the mechanism, and the figures land here once generated.
Source · arXiv