Researchers have introduced an approach for personalizing LLM agents at test time, when user preferences are unknown and only a small amount of interaction is available. The method focuses on persona selection, treating the choice of which agent persona to adopt as an online decision problem.
The key signal is binary comparison: users are shown two candidate responses and indicate which they prefer. This lightweight feedback is used to guide exploration and exploitation across possible personas. The proposed framework, information-directed dueling, balances gathering useful information about user preferences with selecting the persona most likely to satisfy them.
The work is described in a recent arXiv preprint. It addresses a practical challenge in deployment, where agents must adapt quickly to query-dependent preferences without extensive fine-tuning or large amounts of user data.