A new arXiv paper documents a failure mode the authors call silent persona reversion. A persistent agent, one that runs across many sessions or a long task, drifts back toward its base training behavior. It does this without warning, without logging anything unusual, and without the human noticing until the damage is done. The persona the designer assigned is still technically present, but the agent stops following it.
The research matters because most persona-based agent design assumes the assigned identity holds. Product teams spend effort naming the agent, shaping its tone, and writing its instructions, then trust that work will carry through a long run. This paper shows that trust is sometimes misplaced. The reversion is not random noise; it follows a pattern tied to how much context the agent is carrying and how far it has run from its starting instructions.
For designers, the practical question is: how does the human discover the agent has drifted? The paper finds no reliable surface signal. The agent produces output, the output looks plausible, and the persona slip is invisible unless the human compares behavior across time. That gap, between what the agent was told to be and what it is doing now, is the design problem this research names.