Hailong Shi

Long-context LLM agents increasingly serve multiple users or personas within a single session, requiring stable identity and knowledge boundaries under frequent switching. We identify a common failure mode, identity drift, where models conflate user-specific states and leak information across rol...