Qiyao Ma

ShallowDefend: Harnessing Shallow Safety Alignment for Multi-Turn Defender

According to the abstract, Qiyao Ma, Xiaodong Liu, Weiwei Yang and nine co-authors address multi-turn jailbreaks, where a harmful request spread across a conversation still succeeds. They propose training models to assess intent, commit to an explicit safety decision, then respond, using reinforcement learning with a rule-based reward.