ClientCast

Towards a Client-Centered Assessment of LLM Therapists by Client Simulation

Venue: ArXiv
Paper: ArXiv

Overview

ClientCast creates client simulations for assessing LLM-based therapists. It introduces well-designed and detailed psychological profiles including Big Five personality traits, symptom assessments (PHQ-9, GAD-7, OQ-45), and conversation history from real therapy sessions to create realistic client behavior.

Key Features

• Deep Psychological Profiling: Models Big Five personality traits alongside critical therapy-relevant traits like resistance and emotional fluctuation.
• Symptom Assessment: Grounded in validated scales (PHQ-9, GAD-7, OQ-45).
• Real Conversation Grounding: Uses human therapy conversation excerpts
• Client-Centered Focus: Enables the simulated client to evaluate the therapist

How It Works

1. Profile Construction: Builds case synopsis from basic profile
2. Symptom Integration: Incorporates identified symptoms with severity levels
3. Personality Mapping: Uses Big Five traits to guide response style
4. Conversation Grounding: References real therapy conversation excerpts
5. Response Generation: Produces responses consistent with profile and conversation context

Psychological Scales

ClientCast integrates multiple validated assessment instruments:

Scale	Description	Focus
PHQ-9	Patient Health Questionnaire	Depression symptoms
GAD-7	Generalized Anxiety Disorder scale	Anxiety symptoms
OQ-45	Outcome Questionnaire	Overall functioning

Usage

CLI

uv run python -m examples.simulate client=clientCast

Python

from patienthub.clients import get_client

client = get_client(agent_name="clientCast", lang='en')

response = client.generate_response("What brings you in today?")
print(response.content)

Configuration

Option	Description	Default
prompt_path	Path to prompt file	data/prompts/client/clientCast.yaml
data_path	Path to character file	data/characters/ClientCast.json
conv_path	Path to conversation data	data/resources/ClientCast/human_data.json
symptoms_path	Path to symptom assessment definitions	data/resources/ClientCast/symptoms.json
data_idx	Character index	0
conv_id	Conversation excerpt ID	0

Character Data Format

{
  "basic_profile": {
    "name": "Not Specified",
    "gender": "Female",
    "age": -1,
    "occupation": "Not Specified",
    "topic": "alcohol consumption – discussing frequency and effects...",
    "situation": "The client drinks alcohol regularly to unwind...",
    "emotion": "The client feels guilty about their drinking habits...",
    "reasons": "Concerned about alcohol consumption and its effect on depression",
    "problem": "alcohol consumption – struggles with drinking more than intended",
    "feeling_expression": "Medium – shares feelings but does not deeply explore them",
    "emotional_fluctuation": "Medium – talks about relief and worsening feelings",
    "resistance": "Low – open to discussing drinking habits"
  },
  "big_five": {
    "Openness": {
      "score_percent": 55,
      "explanation": "Enjoys watching movies and reading..."
    },
    "Conscientiousness": {
      "score_percent": 60,
      "explanation": "Aware of drinking habits, willing to set goals..."
    },
    "Extraversion": {
      "score_percent": 45,
      "explanation": "Enjoys social activities but doesn't mention frequent interactions"
    },
    "Agreeableness": {
      "score_percent": 65,
      "explanation": "Willing to reduce alcohol consumption, cooperative"
    },
    "Neuroticism": {
      "score_percent": 75,
      "explanation": "Experiences depression and uses alcohol to cope"
    }
  },
  "symptoms": {
    "PHQ-9": {
      "1": {
        "identified": false,
        "severity_level": null,
        "explanation": "Still finds interest in activities..."
      }
    },
    "GAD-7": {
      "1": {
        "identified": false,
        "explanation": "No mentioned anxiety symptoms..."
      }
    },
    "OQ-45": {
      "1": {
        "identified": false,
        "explanation": "Unable to assess social interactions..."
      }
    }
  }
}

Resources

data/resources/ClientCast/human_data.json: Human therapy conversation excerpts for grounding client simulations data/resources/ClientCast/symptoms.json: Symptom item texts for PHQ-9, GAD-7, and OQ-45 used to identify symptoms in the ClientCast profile