Large language models are increasingly employed to simulate clients for counselor training and psychological counseling research. Reliable simulation requires the client to stay psychologically coherent throughout extended dialogues. Existing role‑playing approaches rely on static profile prompts, leading to persona drift, unrealistic cooperativeness, and inconsistent psychological states, communicative actions, and emotions. Moreover, current evaluations lack a unified testbed that captures both stable client characteristics and evolving dynamics.
We introduce MASC (Multi‑Agent Self‑Calibration), a framework that aligns latent constructs to achieve consistent client role‑playing. MASC consists of four core components:
- Construct‑guided generation – uses session‑level profiles and Big‑Five personality traits to steer the model toward psychologically plausible responses.
- Collaborative refinement – multiple agents jointly polish the output, ensuring natural language and counseling relevance.
- Consistency verification – continuously checks for mismatches in personality, emotion, and behavior, flagging potential drifts.
- Memory‑based revision – leverages dialogue memory to retroactively correct inconsistencies, forming a closed‑loop calibration.
To benchmark the framework, we create CRPC‑Bench, which includes:
- Session‑level profile information and Big‑Five personality annotations;
- Turn‑level psychological state, communicative action, and emotion labels. The benchmark comprises 38 motivational interviewing client profiles enriched with personality and emotion annotations.
Experiments demonstrate that MASC outperforms existing methods across profile fidelity, personality consistency, receptivity, and turn‑level consistency, with the heterogeneous configuration achieving the strongest overall performance. MASC and CRPC‑Bench provide a unified foundation for developing and evaluating psychologically coherent client simulations for AI‑assisted counseling research and training.
Review