Can AI simulation teach interpersonal skills more effectively?

Most AI systems for therapeutic training focus on either conversational skills or emotional regulation. IMBUE is the first to address both simultaneously, grounding its approach in the DEAR MAN framework from Dialectical Behavioral Therapy — which includes conversational strategies (Describe, Express, Assert, Reinforce, Negotiate) and a desired state of mind (Mindful, Confident) for productive conversations.

The key technical contribution is a prompting strategy that demonstrates contrasting pairs of strong and weak utterances, outperforming GPT-4 by 24.8% on skill use evaluation and producing more expert-like, specific, and actionable improvement suggestions. Through a randomized trial of 86 participants, the simulation-only variant significantly improved self-efficacy (up to 17%) and reduced negative emotions (up to 25%).

A critical design choice: the system focuses on the wellbeing and mindfulness of the conversation participant rather than optimal negotiation outcomes. "We consider it a suboptimal case if someone 'wins' a negotiation but was not being mindful and had negative emotional swings during the process." This reframes success from task completion to process quality — a rare approach in AI-assisted training systems.

The formative study with psychology experts yielded insights that echo broader patterns in the vault: experts emphasized that effective training requires grounding in clinical theory, not just pattern matching, and that feedback must be specific and actionable rather than generic encouragement. Since the gap between simulating therapy skills and implementing them therapeutically remains unresolved, IMBUE represents a step toward implementation by training users rather than replacing therapists.