💻 computer science

PatientAct: Theory-Grounded Mental Health Client Simulation

PatientAct is a theory-grounded framework for simulating mental health clients that overcomes the limitations of current overly cooperative LLMs by integrating 5Ps clinical case formulations and dynamic memory layers with trust thresholds to produce behaviorally realistic, resistant, and clinically plausible client interactions.

Original authors: Sahand Sabour, TszYam NG, Yaqian Chen, Guanqun Bi, Jialu Zhao, Minlie Huang

Published 2026-08-14

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Sahand Sabour, TszYam NG, Yaqian Chen, Guanqun Bi, Jialu Zhao, Minlie Huang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to be a therapist. You wouldn't just tell it, "Be nice and listen." You'd need to give it a brain that understands human complexity. This is the world of Large Language Models (LLMs), which are super-smart computer programs that can talk like humans. Scientists are using these programs to create simulated clients—digital people that pretend to have emotional problems so real therapists can practice on them without risking a real patient's feelings. But here's the catch: most of these digital patients are too perfect. They are like overly polite guests who tell you their deepest, darkest secrets the moment you walk through the door, agree with everything you say, and magically fix their problems in five minutes. Real humans don't work like that; we get defensive, we hide our pain until we trust you, and we sometimes just don't want to talk about our childhood.

Enter PATIENTACT, a new framework designed to fix these "too-cooperative" digital patients. The researchers built a system that gives simulated clients a much deeper, more realistic backstory and a set of rules for how they share information. Instead of just dumping a whole biography into the computer's memory, PATIENTACT treats the client's mind like a house with locked rooms. Some doors (like "I have trouble sleeping") are unlocked immediately, but others (like "I was bullied as a kid") stay locked until the "therapist" earns enough trust to get the key. By using established psychological theories to map out why a person feels the way they do, the system creates digital clients that can actually say "no," get upset, or change the subject, just like a real person would.

The Problem: The "Yes-Man" Digital Patient

For a long time, scientists have been trying to build these simulated clients to help train new therapists and test out AI therapists. But there was a major glitch. The digital patients were acting like "yes-men." They were too eager to please. If a therapist asked, "How did that make you feel?", the digital client would immediately spill their entire life story, even if they had only just met. They would accept every suggestion without a fight and solve their biggest emotional crises in a single conversation.

The researchers realized this happened because the digital profiles were too shallow. They listed what the client felt (e.g., "I am sad") but didn't explain why they felt that way or how those feelings were built up over a lifetime. Furthermore, the computers treated all information as equally easy to share. In reality, a person might happily talk about their bad sleep schedule but would shut down if you asked about a painful memory from when they were ten. The old systems didn't understand that trust is a currency you have to earn, and they didn't know how to make a client "resist" in a realistic way.

The Solution: A House with Locked Doors

To fix this, the team created PATIENTACT, which stands for a framework grounded in real clinical theories. Think of it as building a digital human with a "psychological skeleton" rather than just a skin.

1. The Deep Backstory (The 5Ps)
Instead of a simple list of symptoms, the system builds a profile using the 5Ps framework. Imagine a detective trying to solve a mystery. They don't just look at the crime scene (the Presenting Problem); they look at who was vulnerable before the crime (Predisposing Factors), what triggered it today (Precipitating Factors), what keeps the problem going (Perpetuating Factors), and what strengths the person has (Protective Factors). This gives the digital client a "causal depth." They don't just say "I'm anxious"; they can explain that their anxiety comes from a specific childhood event and is kept alive by a habit of avoiding social situations.

2. The Trust Gatekeeper
This is the most magical part. The researchers split the client's memory into two layers: a Static Layer (things everyone knows, like their job or name) and a Dynamic Layer (deep secrets).

The Key Mechanism: Every piece of information in the Dynamic Layer has a "trust threshold."
How it works: If the therapist asks about a surface issue, the client answers. But if the therapist asks about a sensitive topic (like a childhood trauma) before the client feels safe, the system checks the trust level. If the trust isn't high enough, the door stays locked.
The Result: The client doesn't just say "I can't tell you." They might get quiet, change the subject, or get defensive. This is called resistance, and it's a huge part of real therapy.

3. The Emotional Reaction Loop
Before the digital client types a single word, the system runs a mini-movie in its head.

Step 1: How does the client feel about what the therapist just said? (Are they challenged? Scared? Understood?)
Step 2: Based on that feeling and their current trust level, what do they do? (Do they open up? Do they shut down? Do they get angry?)
Step 3: Only then does the system generate the actual response.

This means the client's behavior isn't random; it's a chain reaction of feelings and trust. If a therapist pushes too hard, the client might get defensive. If the therapist is gentle, the client might slowly open up.

What They Found: Realism Wins

The team tested their new system against three other popular methods. They created 40 different clinical scenarios (mostly involving depression and anxiety) and had human experts and AI judges rate the conversations.

The results were clear: PATIENTACT was significantly more realistic.

Better Resistance: The digital clients pushed back in varied, natural ways. They didn't just say "no"; they went silent, talked about something else, or got a little snappy, depending on their personality.
Better Pacing: Information was revealed at a natural speed. Secrets weren't dumped immediately; they came out slowly as the "relationship" with the therapist grew.
The "Depth" Factor: The researchers found that having a deep, well-structured backstory (the 5Ps) was even more important than having fancy dynamic rules. A client with a rich history felt more real than a client with a shallow history, even if the shallow one had some dynamic features.

Interestingly, the AI judges sometimes rated the other systems higher on "coherence" (how well the story held together), but the human experts preferred PATIENTACT for its emotional authenticity and realistic resistance. This suggests that while computers can spot a smooth story, humans are better at spotting a real person.

Why This Matters

This isn't about replacing human therapists with robots. It's about giving human therapists better training tools. If a student therapist can practice on a digital client that actually fights back, gets scared, or hides their pain, they will be much better prepared for the real thing. PATIENTACT suggests that to build truly helpful AI for mental health, we need to stop making digital humans that are too eager to please and start building ones that have the messy, guarded, and complex hearts of real people. The paper shows that by grounding these simulations in real psychological theories and respecting the "trust gate," we can create a much more powerful tool for learning and research.

Technical Summary: PATIENTACT

Problem Statement

Current Large Language Model (LLM)-based simulated clients, used for training novice counselors, evaluating AI therapists, and generating synthetic data, suffer from a critical lack of behavioral realism. Existing simulators produce "overly cooperative" clients that disclose sensitive information too readily, accept therapeutic reframes without resistance, and resolve core psychological issues within a single session. The authors attribute these failures to two primary design shortcomings in prior work:

Lack of Causal Depth: Existing profiles describe what a client thinks and feels (symptoms, beliefs) but fail to explain why these states exist (vulnerabilities, triggers, maintenance cycles). Consequently, simulators cannot coherently explain a client's history or exhibit the specific maintenance behaviors that characterize real therapy.
Uniform Accessibility of Content: Prior approaches either expose the entire client profile in the system prompt or apply a single global control (e.g., a static "resistance" label) to all content. This fails to capture the reality that clients selectively share information based on trust levels; for instance, a client might freely discuss sleep issues while actively guarding childhood memories until a therapeutic alliance is established.

Methodology: PATIENTACT Framework

The authors propose PATIENTACT, a theory-grounded framework for client simulation that addresses these shortcomings through a structured profile schema and a dynamic, multi-step simulation pipeline.

1. Profile Schema (The "What")

PATIENTACT structures client profiles using three components to ensure causal depth and modality-agnosticism:

Demographics: Grounds the client's identity (name, age, occupation, etc.).
Problem Formulation (5Ps): Adopting the 5Ps clinical case formulation (Johnstone and Dallos, 2013), the framework explicitly models:
- Presenting Problem: The current issue.
- Precipitating Factors: Specific events triggering the episode.
- Predisposing Factors: Psychological and social vulnerabilities (e.g., childhood events).
- Perpetuating Factors: Maintenance cycles (e.g., avoidance $\to$ relief $\to$ isolation).
- Protective Factors: Strengths and resources.
Psychological Formulation: Integrates cognitive and interpersonal dynamics:
- Cognitive Layer: Intermediate beliefs, automatic thoughts, triggers, and coping patterns (inspired by Beck, 2020).
- Interpersonal Patterns: Core Conflictual Relationship Themes (CCRT) defining how the client relates to the therapist and others (Luborsky and Crits-Christoph, 1998).

2. Simulation Pipeline (The "How")

The simulation operates through a dynamic pipeline that processes each therapist utterance before generating a response:

Profile Decomposition: The profile is split into a Static Layer (always visible, e.g., demographics) and a Dynamic Memory Layer. Items in the dynamic layer are assigned disclosure thresholds (1.0–4.0) based on their vulnerability, grounded in research on therapeutic trust (Srivastava et al., 2025).
Trust-Gated Retrieval: At each turn, the system matches the therapist's input against activation tags. Content is only retrieved if the client's current trust level meets the item's threshold. Items below the threshold are placed on a "blocked list," creating pressure to deflect without revealing the content.
Reaction and Behavior Modeling: Before generating text, the system explicitly models:
1. Emotional Reaction: Selecting a reaction (e.g., challenged, scared) and intensity based on Hill's (1992) taxonomy.
2. Behavior Selection: Choosing a behavior (e.g., resistance, cognitive exploration) consistent with the reaction, trust level, and coping patterns.
3. Resistance Generation: If resistance is selected, the system determines the specific form (quantity, content, or style) based on Otani's (1989) taxonomy, rather than applying a generic refusal.
Trust Dynamics: After each exchange, the trust level is updated (±0.25 or ±0.5) based on the therapist's behavior and the client's attachment style (anxious, avoidant, or disorganized).

Key Contributions

Modality-Agnostic Profile Schema: A structured schema integrating the 5Ps clinical formulation, cognitive layers, and interpersonal patterns, providing the causal depth missing in prior work without tying the design to a single therapeutic modality (e.g., CBT or MI).
Theory-Grounded Simulation Framework: A pipeline that processes therapist utterances through reaction, behavior selection, and trust-gated retrieval. This allows for dynamic disclosure pacing and multi-dimensional resistance (quantity, content, style) rather than static cooperation or single-label resistance.
Comprehensive Evaluation: A rigorous evaluation demonstrating that PATIENTACT produces significantly more realistic client simulations than existing baselines across expert and LLM-based assessments.

Results

The framework was evaluated on 40 clinical situations covering depression and anxiety, comparing PATIENTACT against three baselines: Patient-ψ (static, theory-grounded), AnnaAgent (dynamic emotion, minimal profile), and ConsistentMI (modality-specific, dynamic).

Profile Generation: Expert annotators rated PATIENTACT profiles highly for Clinical Plausibility (4.43/5), Internal Consistency (4.38/5), and Clinical Depth (4.28/5). Annotators correctly identified the underlying attachment styles and core belief themes at rates significantly above chance.
Simulation Performance: PATIENTACT significantly outperformed all baselines in human evaluations across five dimensions: Coherence, Disclosure Pacing, Resistance Quality, Emotional Authenticity, and Behavioral Realism.
- The most significant gains were in Resistance Quality (+0.67 over the best baseline) and Behavioral Realism (+0.63), directly addressing the "over-cooperative" failure mode of prior work.
- Disclosure Pacing also saw substantial improvement (+0.50), indicating clients shared information at a natural, trust-dependent pace.
Ablation Study: Removing the dynamic memory layer (placing the full profile in the prompt) caused the largest drop in performance, particularly in Resistance Quality, confirming that retrieval-based content selection is crucial for preventing over-disclosure.
LLM vs. Human Evaluation: The study found a divergence between LLM judges and human experts. While LLMs rated PATIENTACT highly on realism, they sometimes preferred baselines on coherence and disclosure pacing. The authors note that dimensions where LLMs disagreed with humans (e.g., Resistance Quality) had the lowest human inter-annotator agreement, suggesting automated evaluation is insufficient for assessing clinical appropriateness.

Significance and Claims

The paper claims that profile depth and quality are primary drivers of simulation realism, yielding larger improvements than dynamic behavioral mechanisms alone. By grounding the simulation in established clinical theories (5Ps, CCRT, Hill's process model), PATIENTACT generates diverse profiles with high clinical plausibility that capture the friction and dynamics of real therapeutic dialogue.

The authors position PATIENTACT as a foundational tool to:

Provide scalable, realistic practice environments for training novice counselors.
Serve as standardized, rigorous test cases for evaluating LLM-based therapists.
Generate diverse synthetic therapy data for psychological research.

The authors remain modest regarding generalization, noting that the current evaluation is limited to depression and anxiety, single 15-turn sessions, and English-language contexts. They acknowledge that extending the framework to other disorders, multi-session dynamics, and diverse cultural contexts requires further evaluation. They also emphasize that the system is a research tool for simulation and not a diagnostic tool or a substitute for real clinical interactions.