Skip to content

Client Simulation via Conversational Trajectory Modeling for Both Trainee Practice and Model Evaluation in Mental Health Counseling

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/qiuhuachuan/PsyCLIENT
Area: Dialogue Systems
Keywords: Client Simulation, Conversational Trajectory Modeling, Counselor Training, LLM Agent, Persona Decoupling

TL;DR

PsyCLIENT tackles the overly compliant and behaviorally uniform nature of simulated counseling clients by decoupling static client profiles from dynamic conversational behavior trajectories, constructing the PsyCLIENT-CP benchmark of 120 Chinese clinical profiles and achieving superior authenticity and training utility.

Background & Motivation

In counselor education and human-AI collaborative counseling, providing novice therapists with safe, scalable, and reproducible practice environments is critical for building core clinical competencies. Traditional training predominantly relies on peer role-play or supervised clinical practice. However, peer role-playing often struggles to capture the authentic clinical resistance, nuanced psychological defenses, and unpredictable behavioral shifts seen in real sessions, while practicing directly on real clients introduces significant ethical and safety concerns. With the rapid evolution of large language models (LLMs) in persona simulation and interactive agents, LLM-based simulated clients have emerged as a promising avenue for scalable trainee training, controllable dialogue synthesis, and interactive benchmarking of automated counseling systems. Nonetheless, existing simulators suffer from three key bottlenecks: severe scarcity of public, clinically grounded case profiles, absence of dynamic behavioral trajectory modeling across multi-turn interactions, and a complete dearth of open resources tailored to Chinese sociocultural counseling contexts.

The fundamental tension lies in the nature of therapeutic communication: real psychological counseling is never a unidirectional, obedient question-and-answer exchange. Instead, clients exhibit complex, nonlinear emotional trajectoriesโ€”ranging from tentative disclosure and hesitation to outright resistance, topic avoidance, intellectualized defense, and eventual cognitive reframing. Conversely, existing LLM-based simulators rely almost exclusively on static system prompts. Due to instruction-tuning and human alignment biases, LLMs display a strong tendency toward being overly compliant or sycophantic. They readily yield to novice counselors' gentle questioning, agree to interventions prematurely, and produce unrealistic "instant breakthroughs," thereby creating a distorted training environment that fails to test trainees' abilities to navigate resistance and build genuine therapeutic alliances.

This paper's angle of attack is to explicitly decouple "who the client is" (static clinical persona profile) from "how the client behaves at each turn" (dynamic conversational behavior trajectory). By extracting discrete semantic behavior labels and authentic phrasing patterns from real multi-turn counseling dialogues, an LLM generator can be conditioned to unfold behaviorally guided interactions turn by turn. Core idea: decompose counseling client simulation into the orthogonal modeling of static clinical profiles and dynamic conversational trajectories, using discrete psychological behavior labels and style references under a soft consistency constraint to produce highly authentic, resistance-capable counseling interactions.

Method

Overall Architecture

The PsyCLIENT framework consists of three core components: PsyCLIENT-CP (a dataset of 120 clinically formulated Chinese client profiles spanning 60 counseling topics), a repository of 324 behavior-annotated conversational trajectories extracted from real human-client counseling sessions, and a soft-constrained LLM generation engine. At each turn, the simulator conditions on the static profile, the running dialogue history, the current target behavior label, and reference utterance styling, generating realistic responses while maintaining global character consistency.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Counselor Training Session or Model Benchmark"] --> B["Clinically Grounded Profile Construction<br/>SOAP alignment + 60 topics + 120 detailed profiles"]
    C["Real-World Human Counseling Corpus"] --> D["Discrete Behavior Trajectory Extraction<br/>12 atomic behavior labels + style reference stream"]
    B --> E["Soft-Constrained Trajectory-Conditioned Generation<br/>Profile + History + Target Behavior + Reference Content"]
    D --> E
    E --> F["Orthogonal Profile-Trajectory Decoupled Synthesis<br/>Independent recombination of profiles and trajectories"]
    F --> G["Output: Highly Authentic Multi-Turn Counseling Dialogue<br/>For trainee practice and interactive LLM evaluation"]

Key Designs

1. Clinically Grounded Profile Construction: Decoupled and High-Density Case Representations Existing client role-play approaches typically rely on brief, superficial persona snippets (e.g., age, job, and a single presenting complaint), failing to capture the multidimensional psychological depth required in clinical training. To bridge this gap, PsyCLIENT synthesizes four major clinical documentation frameworks: SOAP notes (Subjective, Objective, Assessment, Plan), Intake Session forms, formal Counseling Reports, and CBT Case Conceptualizations. Collaborating with licensed counseling psychologists across 60 representative topics (such as mate selection, family conflicts, career burnout, suicidal ideation, and psychosomatic complaints), the authors developed PsyCLIENT-CP containing 120 detailed Chinese client profiles. Each profile systematically covers basic demographics, somatic/psychiatric diagnoses, personality traits, physical appearance, hobbies, lifestyle habits, developmental milestones, significant relational traumas, recent triggers, and specific communication styles (e.g., guarded/reserved or overly verbose). With an average length of 3,930 characters (74% exceeding 2,000 characters), these profiles provide dense factual anchors and psychological boundaries that prevent out-of-character hallucination.

2. Discrete Behavior Trajectory Extraction: Capturing Dynamic Client Reactions from Real Sessions To prevent the LLM from degenerating into uniform compliance, PsyCLIENT formalizes client conversational acts across a discrete semantic space \(\mathcal{A}\) containing 12 atomic client behavior labels: $\(\mathcal{A} = \{\text{co, pi, rr, ex, re, ec, de, sh, st, fd, sa, ot}\}\)$ These comprise Confirming (co), Providing Information (pi), Reasonable Request (rr), Extending (ex), Reformulating (re), Expressing Confusion (ec), Defending (de), Self-criticism or Hopelessness (sh), Shifting Topics (st), Focus Disconnection (fd), Sarcastic Answer (sa), and Other (ot). The authors filtered real-life Chinese counseling transcripts to retain sessions exceeding 30 turns, extracting 324 conversational trajectories \(\tau = \{(B_t, D_t)\}_{t=1}^T\). Here, \(B_t \subseteq \mathcal{A}\) designates the target behavior set at turn \(t\), while \(D_t\) is the de-identified natural utterance from the human client, serving as a reference for phrasing length, lexical informalities, and authentic hesitation patterns.

3. Soft-Constrained Trajectory-Conditioned Generation: Balancing Behavior Adherence with Dialogue Coherence Blindly forcing an LLM to follow a pre-assigned behavior label can break conversational flow when the counselor asks a direct factual question that clashes with defensive maneuvers (e.g., asking for basic marital status when the assigned label is Defending). To address this rigidity, PsyCLIENT incorporates a "coherence-first soft constraint" within its prompt architecture: the generator is instructed that if the counselor's intervention strictly necessitates a specific logical response, contextual coherence takes priority over rigid label execution, and the deviation is internally noted. Conditioned on client profile \(x\), history \(h_t\), target behavior \(B_t\), and reference style \(D_t\), the simulator produces responses that naturally exhibit defensiveness and skepticism without feeling disjointed or robotic.

4. Orthogonal Profile-Trajectory Decoupled Synthesis: Enabling Scalable and Combinatorial Simulation In training and evaluation scenarios, tightly coupling a behavior trajectory to a single case profile restricts diversity. The paper establishes the theoretical Universal Realizability: under standard linguistic expressivity and counselor policy executability assumptions, any finite-length trajectory \(\tau_B\) can be realized across arbitrary client profiles \(x\). In validation studies with licensed counseling experts, evaluating both "fixed trajectory across varying profiles" and "fixed profile across varying trajectories," experts unanimously awarded maximum realism scores (7.0 / 7.0 on a 7-point Likert scale). This empirical and theoretical independence allows \(120 \text{ profiles} \times 324 \text{ trajectories}\) to be combinatorially re-paired into over 38,000 diverse, long-horizon simulated counseling sessions.

A Worked Example

Consider a session with Ms. Lin (Profile: 31-year-old high school English teacher struggling with intimacy avoidance and deep-seated distrust of men due to parental infidelity), paired with an extracted trajectory containing defensive and avoidant turns: - Turn 1 (Target Behavior: Providing Information): The novice counselor opens with an inquiry about her week. Ms. Lin responds cautiously and informally: "Actually, I've never been in a romantic relationship before. Lately, due to my age and pressure from my family, I've been feeling very anxious..." - Turn 4 (Target Behavior: Defending): When the counselor attempts to quickly label and analyze her tendency to avoid men, instead of passively concurring, Ms. Lin activates an authentic defense: "I think keeping my distance is the right thing. After all, many men nowadays are just not trustworthyโ€”I've seen that firsthand in my family." - Turn 7 (Target Behavior: Shifting Topics): When pressed gently on painful emotional triggers, she avoids vulnerability: "Today during class, a male colleague came to borrow some materials, and I immediately made an excuse to go to the office." - Turn 12 (Target Behavior: Reformulating): Following empathetic reflection from the counselor, she engages in genuine cognitive reframing: "Maybe it is because I'm afraid of failing that I chose to avoid them. I probably should try to gradually engage with my roommates." Across 30+ turns, the novice counselor experiences genuine therapeutic friction, learning how to de-escalate defenses rather than coasting through artificial consensus.

Loss & Training

PsyCLIENT employs Claude 3.5 Sonnet as its backbone generation engine under standard inference decoding without requiring task-specific parameter fine-tuning. The prompt structure enforces strict persona boundary adherence (e.g., forbidding clients without children from inventing parental anecdotes) and explicitly suppresses the generation of behavior label meta-text. This enables zero-shot, prompt-conditioned behavioral steering that reliably sustains multi-turn interactions exceeding 40 turns without persona collapse.

Key Experimental Results

Main Results

Evaluation was conducted with 24 licensed counseling professionals (mean age = 25.42, 14 females, averaging 356 hours of direct clinical practice) in a double-blind setup across five Authenticity dimensions and five Counselor Training Utility dimensions (1โ€“7 Likert scale).

Client Simulation System Fluency (โ†‘) Emotion (โ†‘) Coherence (โ†‘) Contextual Appropriateness (โ†‘) Overall Authenticity (โ†‘)
AnnaAgent (ACL 2025) 6.00 6.25 4.75 5.25 4.75
ConsistentMI (ACL 2025) 4.50 5.00 3.50 3.75 3.50
Patient-ฮจ (EMNLP 2024) 6.00 4.75 5.25 5.25 4.25
Roleplay-doh (EMNLP 2024) 4.75 5.00 4.50 5.00 4.75
PsyCLIENT (Ours) 6.25 6.25 5.75 6.00 6.00

In the evaluation of perceived training utility by professional counselors:

Client Simulation System Listening (โ†‘) Questioning (โ†‘) Emotion-handling (โ†‘) Technique-practice (โ†‘) Recommendation (โ†‘)
AnnaAgent 5.25 5.00 5.75 5.00 5.25
ConsistentMI 4.25 4.00 3.75 3.75 3.50
Patient-ฮจ 4.25 4.75 5.00 4.00 4.50
Roleplay-doh 5.50 5.00 5.25 4.50 4.50
PsyCLIENT (Ours) 5.75 5.75 5.75 6.00 6.00

Ablation Study

To isolate the contributions of discrete behavior labels and reference content styling, 20 professional counselors evaluated four system variants: profile-only, +content, +behavior, and the full PsyCLIENT.

Ablation Config Fluency (โ†‘) Emotion (โ†‘) Coherence (โ†‘) Contextual Appropriateness (โ†‘) Overall Authenticity (โ†‘)
profile (baseline) 4.37 ยฑ 1.07 4.57 ยฑ 1.25 4.61 ยฑ 1.13 4.59 ยฑ 1.17 4.50 ยฑ 1.39
+content 4.96 ยฑ 1.11 4.98 ยฑ 0.83 4.76 ยฑ 0.90 4.96 ยฑ 0.89 4.82 ยฑ 1.06
+behavior 5.32 ยฑ 0.87 5.07 ยฑ 1.12 5.09 ยฑ 1.22 5.16 ยฑ 1.09 5.11 ยฑ 1.11
PsyCLIENT (Full model) 5.69 ยฑ 0.75 5.80 ยฑ 0.83 5.38 ยฑ 0.95 5.56 ยฑ 0.91 5.64 ยฑ 0.79

Additionally, in 30-turn Source Discrimination tests, five state-of-the-art LLM judges (Qwen3.6-max, Kimi-k2, MiniMax-M2.7, DeepSeek-V4-Flash, DeepSeek-V4-Pro) evaluated whether dialogues originated from human clients or AI simulators (lower detection numbers reflect higher human-likeness):

Judge Model Human-Client (Acc / 45) profile (Acc / 45) +content (Acc / 45) +behavior (Acc / 45) PsyCLIENT (Acc / 45)
Qwen3.6-max-preview 43 / 44 45 / 45 42 / 43 44 / 41 40 / 37
Kimi-k2-0905-preview 39 / 45 41 / 32 38 / 24 41 / 22 40 / 7
MiniMax-M2.7 32 / 45 32 / 5 32 / 5 27 / 8 22 / 3
DeepSeek-V4-Flash 34 / 38 45 / 19 42 / 25 41 / 22 38 / 16
DeepSeek-V4-Pro 42 / 41 37 / 41 36 / 32 32 / 36 32 / 31
Total Detection Rate 89.6% (403/450) 76.0% (342/450) 70.9% (319/450) 69.8% (314/450) 59.1% (266/450)

Key Findings

  • Behavior Conditioning Drives the Primary Authenticity Jump: Adding reference content (+content) yields modest improvements (Overall Authenticity +0.32 over profile-only), whereas explicit behavior conditioning (+behavior) drives a substantial +0.61 gain, demonstrating that structured behavioral steering is the primary driver in eliminating artificial sycophancy.
  • Exceptional Human Expert Indistinguishability: In blinded human discrimination evaluations, experts correctly classified 100% (10/10) of dialogues from AnnaAgent, Patient-ฮจ, and ConsistentMI as AI-generated. For PsyCLIENT, 2 out of 10 complete 30-turn dialogues were misclassified as real human clients. In broader multi-sample tasks, expert AI-detection rates for PsyCLIENT dropped to 0.048โ€“0.067.
  • Robustness in Extended Interactions: Across automated LLM evaluations on 30-turn sessions, the discrimination accuracy for PsyCLIENT fell to 59.1% (compared to 76.0% for the profile baseline and 89.6% for human controls), confirming that trajectory modeling successfully prevents persona decay over extended interactions.

Highlights & Insights

  • Decoupled Architecture Solves Data Scarcity and Privacy: By decoupling clinical profiles from behavioral trajectories, the framework circumvents the privacy obstacles of releasing raw patient records while multiplying simulation scenarios through orthogonal recombination.
  • Targeting Client Resistance to Cure Model Sycophancy: Explicitly incorporating defense, topic shifting, and sarcastic reaction labels into dialogue planning breaks the intrinsic sycophantic bias of instruction-tuned LLMs.
  • Soft Constraints Prevent Rigid Overfitting: Balancing trajectory adherence with context coherence allows simulated clients to navigate natural conversational tangents while preserving overall clinical realism.

Limitations & Future Work

  • Single-Session Scope: Current modeling focuses on single sessions (up to 40 turns); real-world psychotherapy typically spans multiple sessions across months, requiring long-term state evolution, therapeutic rupture-and-repair modeling, and cross-session memory mechanisms.
  • Sociocultural Specificity: The PsyCLIENT-CP profiles are deeply grounded in Chinese sociocultural realities (e.g., gaokao pressures, parental marriage expectations), which may require cultural adaptation before deploying in other clinical training environments.
  • Static Trajectory Replay: Behavior trajectories are currently replayed from human transcripts; future work could explore dynamic trajectory planning agents that adaptively select behavior labels in real-time response to trainee interventions.
  • vs Patient-ฮจ (EMNLP 2024): Patient-ฮจ incorporates cognitive behavioral therapy models into prompts but lacks fine-grained behavioral trajectory steering, resulting in lower perceived authenticity (4.25 vs 6.00 for PsyCLIENT) and an inability to challenge trainees with realistic defensiveness.
  • vs ConsistentMI (ACL 2025): ConsistentMI models client mental states via explicit finite-state transitions in motivational interviewing, but its rigid state transitions often degrade conversational fluency and emotional spontaneity; PsyCLIENT's decoupled trajectory and reference-styling paradigm achieves superior language naturalness.

Rating

  • Novelty: โญโญโญโญโญ [Pioneering decoupling of static clinical profiles and dynamic discrete behavior trajectories for client simulation]
  • Experimental Thoroughness: โญโญโญโญโญ [Extensive double-blind evaluations by 24 professional counselors and Turing-style source discrimination across 5 advanced LLMs]
  • Writing Quality: โญโญโญโญโญ [Clear mathematical formalization, comprehensive clinical grounding, and rich case illustrations]
  • Value: โญโญโญโญโญ [Provides the first large-scale, open Chinese clinical client simulation dataset and a robust benchmark for counselor training and dialogue system evaluation]