Study: LLM agents that adapt to user personality boost satisfaction but cut truthfulness

A new arXiv paper from Maryam Shoaeinaeini, Brent Harrison, and A.B. Siddique tackles a key question: can large language models (LLMs) express and adapt to personality in goal-directed conversations without losing task performance? The researchers propose a training-free framework that simulates a task-oriented dialogue (TOD) between two LLMs—a user agent exhibiting a target persona and a system agent that must complete the task while adapting to that persona.

They evaluated three state-of-the-art models—GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash—on Hotel and Restaurant dialogues from the Schema-Guided Dialogue (SGD) dataset. The system agent operated under three conditions: Neutral (no persona info), Try (infers persona from dialogue cues), and Oracle (persona explicitly provided). The user agent was prompted to express all Big Five traits and their opposites, and the researchers measured task success via constraint satisfaction, inform rate, user satisfaction, and truthfulness.

Results show a clear trade-off: adapting to the user's personality improves constraint satisfaction, inform rate, and user satisfaction, but reduces truthfulness. Oracle gains grew when the target trait was strongly expressed, while Try's gains were largely insensitive to realization strength—meaning cue-based adaptation is more robust. Overall, the Try condition best resolves the personalization-task grounding trade-off, offering a reliable route to personality-aware TOD without any fine-tuning. The study also notes that some personality traits are far less reliably expressed by the user agent, highlighting a limitation in current LLM persona control.