Tag
This paper proposes Dialogue Critic Guided Sampling (DCGS), a framework that defends LLMs against multi-turn adversarial attacks by inferring user intent from conversation history and using value/regret-based critics to score responses, achieving improved robustness without fine-tuning.