A new framework called MASC aims to keep AI-simulated therapy clients from breaking character mid-session.
Researchers behind an arXiv paper introduce MASC, a Multi-Agent Self-Calibration framework with latent construct alignment for consistent client role-playing in psychological counseling training. It combines construct-guided generation, collaborative refinement, consistency verification, and memory-based revision in a closed calibration loop that catches and corrects inconsistencies as a dialogue unfolds. The team also built CRPC-Bench, a benchmark of 38 motivational-interviewing client profiles annotated with Big-Five personality traits, turn-level psychological states, communicative actions, and emotions. In testing, MASC beat existing role-playing methods on profile, personality, receptivity, and turn-level consistency, with a heterogeneous agent configuration performing best.
Counselor training often relies on scripted actors or static chatbot personas that either agree with everything the trainee says or lose their personality over a long conversation, which makes for weak practice reps. A simulated client that holds a consistent psychological profile across many turns is a precondition for using AI in clinical training at scale, and the benchmark itself may prove more useful than the framework, since it is one of the first to measure turn-by-turn consistency rather than just surface plausibility.
Worth noting: this is a research prototype tested against other AI methods, not against real training outcomes with actual counselors, so whether more consistent personas produce better-trained therapists is still an open question.