A good advisor changes approach: comforting someone in crisis, challenging someone in denial, and giving practical guidance for a logistical question. A new paper on arXiv (2607.08326) asks whether large language models do the same, or instead fall into a fixed pattern the authors call "persona collapse."

The paper defines five advisory styles and investigates whether LLMs shift between them as the situation demands. The abstract is cut off before listing the five, but the framing is clear: the authors are concerned with situational adaptability in AI-generated advice, not just factual correctness.

They also propose ways to diagnose when a model is stuck in a single persona and to repair that behavior. The full details of the diagnostic and repair methods are not visible in the abstract, but the paper's direction points toward making LLM advisors more context-sensitive and human-like in their responses.