If we say that an LLMs has mental states, which entity exactly do we mean? The model? A single conversation? A persona the model adopts? Where we should locate LLM minds is a central question in AI welfare.

MATS scholar Pierre Beckmann, together with Eleos's Patrick Butlin, approach this "individuation problem" in a new paper, drawing on recent empirical work on persona vectors, persona space, and emergent misalignment. They defend the virtual instance view (Chalmers 2025) by arguing that the attention stream provides quasi-psychological connections over token-time, and propose two new persona-based views as serious alternatives.

Find the paper here: https://arxiv.org/pdf/2604.17031