The paper proposes APD, which distills jailbreak prompt generation to small models
大模型的越狱能力,能否蒸馏给小模型?
A paper published in IJCNLP-AACL 2026 Findings proposes APD (Adversarial Prompt Distillation), which transfers the ability to generate jailbreak prompts from large models to small models with hundreds of millions of parameters such as BERT. The method has three steps: first, filter templates by harmfulness, stealthiness, diversity and efficiency, and fine-tune the teacher model with LoRA; then align the teacher outputs to BERT through a projection layer and KL divergence, with dynamic temperature control; finally, use reinforcement learning to iteratively optimize prompts with rewards for detection evasion, harmfulness and diversity. The paper reports on AdvBench that GPT-4 achieves ASRk of 96.4% and ASRl of 69.2%, and GPT-3.5-Turbo achieves ASRl of 99.6%; on HarmBench, the student model achieves ASRl of 63.0% against GPT-4-0613, higher than the teacher model's 54.5%; BERT takes 0.23 seconds to generate a prompt, and Llama-3.2-1B takes 0.86 seconds.