PsyCLIENT: Virtual Client Simulation Framework

• PsyCLIENT is a simulation framework that generates diverse, authentic virtual clients for mental health counseling training using behavior trajectory modeling.
• It overcomes traditional limitations by employing explicit behavioral conditioning and richly parameterized profiles to ensure dynamic, realistic interactions.
• Empirical evaluations demonstrate high authenticity and improved training effectiveness across various linguistic and cultural contexts.

PsyCLIENT is a simulation framework for generating and evaluating virtual clients in mental health counseling, optimized to produce diverse, authentic, and behaviorally realistic interactions for counselor training and system assessment. By leveraging conversational trajectory modeling, explicit behavioral conditioning, and richly parameterized client profiles, PsyCLIENT addresses longstanding technical and methodological challenges in the automated simulation of counseling dialogues, including profile diversity, dynamic behavior realism, and effective deployment across global linguistic contexts (Qiu et al., 12 Jan 2026).

1. Limitations of Prior Client Simulation Approaches

Traditional LLM-based client simulators exhibit three core limitations:

• Profile Diversity and Realism: Existing datasets are highly constrained—often English-only, synthetic, and lacking in both topic breadth and psychological complexity due to privacy and ethical restrictions. This leads to a narrow range of repeatable client archetypes, preventing robust evaluation of counselor adaptability and system generalizability (Qiu et al., 12 Jan 2026).
• Behavioral Modeling Frameworks: Previous simulators typically focus on compliant client behavior and neglect resistant or ambivalent behaviors typical in real therapeutic exchanges. Over-acquiescent simulated clients result in interactions that do not stress-test the full range of counseling skills or reflect naturalistic conversational trajectories (Qiu et al., 12 Jan 2026).
• Language and Cultural Scope: Almost all published frameworks focus on English, leaving critical gaps in counselor training and system evaluation for non-English and especially Chinese-language clinical environments (Qiu et al., 12 Jan 2026).

PsyCLIENT is architected to overcome these challenges through data-sourced client profiles, explicit behavior trajectory annotation, and culture-sensitive profile banks.

2. Conversational Trajectory Modeling and Behavioral Conditioning

PsyCLIENT relies on a trajectory-based framework wherein simulated dialogs are generated according to predefined behavioral flows:

• Atomic Behavior Labels: Each client turn is indexed by a set of behavior labels from a closed vocabulary of 12 types, including Confirming, Giving Information, Expressing Confusion, Defending, Shifting Topic, and others. Formally, for each turn $t$, the behavior $B_t \subseteq A$ (with $A$ as the behavior set) guides generation.
• Trajectories: The flow $\mathcal{F} = (B_1, ..., B_T)$ specifies the intended behavioral dynamics across the session, extracted and annotated from real counseling data.
• Conditioned Generation: At each turn, the client utterance $u_t = G(x, B_t)$ is produced by an LLM conditioned on psychological profile $x$, explicit behavior label $B_t$, and the preceding dialog context. The system enforces “soft constraints,” allowing logical override of prescribed behavior to preserve coherence regarding the counselor’s intervention style (Qiu et al., 12 Jan 2026).
• Counselor Strategy and Realizability: There is a mapping $m: P(A) \to P(S)$ between possible client behaviors and admissible counselor strategies $S$ (e.g., Reflection, Open Question), enabling targeted simulation of skill-driven response patterns. Universal realizability is guaranteed under expressivity and executability assumptions.

This approach enables simulation of compliant, resistant, and mixed behavior trajectories, supporting both challenging and supportive training scenarios.

3. Client Profile Parameterization and PsyCLIENT-CP Dataset

PsyCLIENT introduces high-fidelity client profiles and large-scale behavioral data resources:

• Profile Structure: Each profile is grounded in real intake forms, SOAP notes, and case conceptualizations, comprising multidimensional attributes (e.g., presenting problem, history, emotion characteristics, coping resources) (Qiu et al., 12 Jan 2026).
• Dataset Scale and Coverage: PsyCLIENT-CP, the companion dataset, contains 120 Chinese-language client profiles spanning 60 counseling topics (e.g., family conflict, career stress, self-harm ideation). Profiles average nearly 4,000 characters and are matched with 324 behavior-annotated conversational trajectories, resulting in 38,880 distinct simulated client/trajectory combinations (Qiu et al., 12 Jan 2026).
• Behavioral Annotation: All trajectories are manually labeled turn-by-turn, facilitating precise conditioning and diverse scenario generation.

This comprehensive resource enables systematic coverage of psychological phenomena—addressing the scarcity of linguistically and clinically rich profile corpora.

4. Prompting, Conditioning, and Utterance Generation Mechanisms

PsyCLIENT’s simulation pipeline employs carefully engineered in-context prompting without additional fine-tuning:

• Prompt Variants: Four template regimes are used: vanilla (profile only), +content (with style exemplar), +behavior (with behavior label), and the full PsyCLIENT prompt (profile + behavior label + style exemplar with soft logical override) (Qiu et al., 12 Jan 2026).
• Response Fidelity: The PsyCLIENT template directs the LLM to generate utterances strictly aligned with the profile and prescribed behavior, matching style, and maintaining interactional appropriateness across turns.
• Content/Coherence Control: If a counselor prompt logically necessitates a behavioral deviation, the generation process prioritizes conversational coherence over rigid adherence to trajectory, yielding lifelike, contextually adaptive simulated clients (Qiu et al., 12 Jan 2026).
• No Fine-tuning: All generation is performed using in-context learning, enabling instant adaptation to new trajectories and profiles.

This mechanism bridges the gap between fixed behavior scripts and the need for realistic, adaptable dialogue.

5. Evaluation Protocols and Quantitative Results

Empirical validation of PsyCLIENT utilizes multi-method evaluation with professional counselor raters and discrimination studies:

• Authenticity and Training Effectiveness: Twenty licensed counselors rated sessions across fluency, emotional expressiveness, coherence, appropriateness, and overall authenticity. PsyCLIENT achieved a mean overall authenticity score of $5.64 \pm 0.79$ (0–7 scale), surpassing all baseline prompt variants (statistically significant, $p<0.01$).
• Discrimination by Experts and LLMs: In blinded discrimination tasks, counselor experts correctly distinguished PsyCLIENT sessions from human–human dialogs in only $\sim$5% of cases (mean confusion rate $\sim$95%). Advanced LLMs (e.g., GPT-4o, Qwen, Claude-S) similarly failed to identify PsyCLIENT-simulated turns (Qiu et al., 12 Jan 2026).
• Training Effectiveness: Across five dimensions (Listening, Questioning, Emotion-handling, Technique-practice, Recommendation), PsyCLIENT outperformed baselines on all ($p<0.01$).
• Fidelity to Profile and Behavior: Qualitative transcript analysis confirms that simulated clients track prescribed behavior trajectories (e.g., alternating between confirmation and resistance) while maintaining persona, symptom, and motivational consistency.

Client Setting	Fluency	Emotion	Coherence	Appropriateness	Overall Authenticity
vanilla	4.37	4.57	4.61	4.59	4.50
+content	4.96	4.98	4.76	4.96	4.82
+behavior	5.32	5.07	5.09	5.16	5.11
PsyCLIENT	5.69	5.80	5.38	5.56	5.64

This table illustrates that simultaneous conditioning on behavior and content in PsyCLIENT’s strategy is critically effective.

6. Comparison to Related Simulation Frameworks

PsyCLIENT advances the client simulation state of the art by integrating principles from prior frameworks and extending their scope:

• Consistent Client Simulation for MI (2025): Defines a modular architecture with formal state tracking ($$S = S_{\mathrm{stage}} \times S_{\mathrm{belief}} \times S_{\mathrm{plan}} \times R$$), state-driven controller, dual-distribution action selection, and LLM-based generation (Yang et al., 5 Feb 2025). PsyCLIENT generalizes beyond motivational interviewing, enabling arbitrary behavioral sequences and adaptation across counseling paradigms.
• ClientCAST (2024): Adopts LLM-based simulated clients for standardized therapist assessment, focusing on client-centered metric extraction (session outcome, alliance, emotion) and profile/trait consistency (Wang et al., 2024). While ClientCAST emphasizes evaluation, PsyCLIENT’s design emphasizes both counselor training and simulation realism via granular behavior control and profile diversity.

A plausible implication is that the combination of explicit state modeling (as in (Yang et al., 5 Feb 2025)) and trajectory conditioning (as in (Qiu et al., 12 Jan 2026)) represents the current methodological standard for client simulation.

7. Limitations, Insights, and Future Directions

Key identified limitations and recommendations include:

• Session Scope: PsyCLIENT currently simulates dynamics within single-session boundaries. Clinical realism would benefit from modeling cross-session (longitudinal) client trajectories and the emergence of phenomena such as inter-session relapse (Qiu et al., 12 Jan 2026).
• Cultural Adaptation: The present profile bank is Chinese-context specific; expansion to other languages and cultures is needed to enhance training applicability for global counseling populations (Qiu et al., 12 Jan 2026).
• Behavioral Adaptation Automation: Automatic, real-time prediction of behavior labels would facilitate dynamic, context-sensitive client modeling and further realism (Qiu et al., 12 Jan 2026).
• Integration with Feedback: Linking simulation to automated feedback pipelines can create end-to-end, data-driven counselor education platforms.

Prompt sensitivity and reliance on in-context learning, rather than fine-tuning, leave current instantiations sensitive to errors accumulating across multi-step prompts. Improved prompt engineering or targeted model adaptation represents a next research frontier (Yang et al., 5 Feb 2025).

In summary, PsyCLIENT’s trajectory-conditioned, profile-rich simulation framework marks a critical advance in the realism, diversity, and training effectiveness of virtual clients for mental health counseling, providing a scalable, research-grade platform for education, evaluation, and cross-cultural benchmarking (Qiu et al., 12 Jan 2026, Yang et al., 5 Feb 2025, Wang et al., 2024).