Everything in this research is simulated and judged by LLMs. It might be hard to prove which of those LLMs struggles with exactly what. The grounding this had

Everything in this research is simulated and judged by LLMs. It might be hard to prove which of those LLMs struggles with exactly what. The grounding this had was that texts produced by role-playing humans (not even actual patients) were closer to texts produced by the patient simulations prompt they decided to end up with than others they tried.