New Framework Boosts LLM Safety Against Multi-Turn Attacks

Large language models face a significant challenge in distinguishing between genuinely harmful user intent and misunderstood but benign queries, especially in multi-turn dialogues where malicious intent can unfold gradually. Current safety mechanisms often treat each turn in isolation, failing to capture the conversational trajectory. Researchers have developed Dialogue Critic Guided Sampling (DCGS) to address this by continuously inferring user intent throughout a conversation. DCGS frames adversarial dialogue as a Markov Decision Process, employing critics at both token and utterance levels to evaluate potential responses. This inference-time reweighting mechanism approximates exponential tilting, ensuring improved expected returns for candidate responses. The framework has demonstrated superior robustness against various adversarial dialogue benchmarks, even transferring its benefits to frontier models without requiring additional fine-tuning.