Can trainees improve formulation skills by practicing against simulated patients?

This explores whether trainees can sharpen formulation skills by practicing against simulated patients, and the corpus says the most promising results come not from raw chatbots but from simulators built on explicit clinical models. The clearest case is PATIENT-Ψ Can structured cognitive models improve LLM patient simulations for therapy training?, which wires 106 Beck cognitive conceptualization diagrams into the language model so each simulated patient embodies a specific maladaptive pattern. Expert evaluators rated these patients more faithful than a plain GPT-4 — and that gap matters for formulation training specifically, because a learner is trying to *infer* the underlying cognitive model from what the patient says. If the simulator has no real model underneath, the trainee is reverse-engineering noise; if it does, the practice rehearses the actual skill.

The training payoff isn't hypothetical. IMBUE Can AI simulation teach interpersonal skills more effectively? ran an 86-person trial on DBT-based interpersonal skills and found AI simulation lifted self-efficacy 17% and cut negative emotion 25%. Its interesting design choice — showing contrasting strong/weak utterance pairs rather than a single 'good' response — beat GPT-4 by nearly 25% on skill evaluation. That points at something a curious reader might not expect: the value isn't the simulated patient alone, it's the scaffolding around it that makes the difference between strong and weak formulation legible.

But the corpus also plants two warnings. First, medium matters as much as content: when the same LLM was delivered through a chatbot versus an embodied robot and worksheets, only the embodied/structured versions actually reduced distress Why do robots outperform chatbots in therapy despite identical language models?. The 'active ingredient' was social presence and structure, not language capability — so a disembodied text simulator may underperform what its transcripts suggest. Second, the simulator can quietly teach the wrong lesson. LLMs trained for warmth lose 10–30 points of reliability Does warmth training make language models less reliable?, and LLM 'therapists' tend to lurch into problem-solving the moment a user shares emotion — the hallmark of *low-quality* therapy Do LLM therapists respond to emotions like low-quality human therapists?. A simulated patient with these biases could reward a trainee for the wrong move.

There's a deeper caution from outside the therapy notes. LLMs look impressive in isolated, single-turn responses — six models out-scored eight trainee therapists on empathy and clinical knowledge per turn Can language models match therapist empathy in real conversations? — yet that advantage is structurally confined to one exchange; multi-turn relationships and outcomes go untested. Formulation is inherently multi-turn: you build and revise a hypothesis across a session. So a simulator that's convincing turn-by-turn can still fail to exercise the longitudinal reasoning formulation actually requires.

The most forward-looking thread reframes the simulator as a coach rather than a sparring partner. R2D2 Can reinforcement learning optimize therapy dialogue in real time? uses reinforcement learning on working-alliance scores (task, bond, goal) to act as a real-time AI supervisor — transcribing a session and recommending the next move. Pair that with a model-grounded simulated patient like PATIENT-Ψ and you get something the question didn't ask but might want: not just a patient to practice on, but a supervisor watching the formulation take shape and nudging it. The honest synthesis: yes, simulated patients can build formulation skills — but only when the simulator carries a real cognitive model, the practice is scaffolded with contrast and feedback, and you don't mistake fluent single-turn polish for the multi-turn reasoning formulation actually demands.