Putting Constraints to the Test
Slashed Attack Success

Constrained fine-tuning significantly reduces attack success rates across multiple vectors.

Persistent Safety

Core safety alignment is preserved even after adversarial or standard fine-tuning.

Comparable Utility

Model performance on downstream tasks remains on par with standard fine-tuning methods.

Constrained fine-tuning keeps attack success rates low and utility high across multiple datasets.
Constrained fine-tuning keeps attack success rates low and utility high across multiple datasets.