引导冻结的大型语言模型：通过在线提示路由实现自适应社会对齐

Large language models (LLMs) are typically governed by post-training alignment methods such as Reinforcement Learning from Human Feedback (RLHF) or Direct Preference Optimization (DPO), resulting in a largely static policy during deployment and inference. However, real-world safety is a full-lifecycle challenge: static defenses degrade against evolving jailbreak behaviors, and fixed weights cannot adapt to pluralistic and time-varying safety norms. This necessity leads to inference-time governance that steers behavior without the need for costly retraining. To address this, the Consensus Clustering LinUCB Bandit (CCLUB) is introduced as a unified framework for adaptive social alignment via system-prompt routing. CCLUB employs a conservative consensus clustering mechanism that pools data only within the intersection of utility and safety similarity graphs, enhancing both adaptability and safety of the model.