Donghua Zhang

The paper proposes a novel safety fine-tuning method that uses the target model's own rollouts to identify and train on the hardest prompts, significantly reducing jailbreak success rates while maintaining usability.