Adaptive and Explicit Safe: Triggering the latent safety awareness of large reasoning models
Adaptive and Explicit safe: Triggering Latent Safety Awareness in Large Reasoning Models
AI summary
The study proposes Safe Trigger, which uses SFT to induce safety labels to trigger the latent safety awareness of large reasoning models, and then uses DPO to improve the correctness and stability of safety analysis. DeepSeek-R1-Distill-Llama-8B reduces the attack success rate on harmful queries and jailbreak benchmarks by an average of 24.65% and 36.72%, respectively, with almost no impact on general performance and user experience.