Multi-user LLM agents

The researchers formalize the interaction of multiple users with LLM agents as a multi-principal decision problem, which the authors describe as the first systematic study in this direction. The paper proposes a unified multi-user interaction protocol and designs three types of stress-test scenarios—instruction following, privacy protection, and coordination—to evaluate current LLMs. Results show that frontier LLMs often fail to maintain a stable priority ordering when user goals conflict, privacy violations gradually increase over multi-turn interactions, and efficiency bottlenecks arise when coordination requires iteratively gathering information.