CARS

When Clients Stop Following: A Cognitive Conceptualization Diagram-driven Framework for Strategic Counseling

Paper: arXiv

Expands a Cognitive Conceptualization Diagram (CCD) and a few seed statements into a full CARS client profile — a persona plus a session-specific cognitive/resistance pattern — following the paper's diverse-profile generation (§3.1.3).

Overview

Property	Value
Key	cars
Type	LLM-based
Output	CARS character files

Key Features

• Persona generation: From a main CCD and seed sentences, produces demographics, family background, interpersonal relationships, physical condition, lifestyle, and chief complaint.
• Session cognitive-pattern generation: From the persona, CCD, and a dialogue excerpt, produces the counseling background, a session-specific CCD (automatic thought, intermediate belief, resistance triggers), preferred counselor style, and other client characteristics.
• Diverse profiles: Different seeds yield distinct cognitive patterns and resistance triggers, matching the paper's goal of a varied evaluation set.

How It Works

generate_character(seed) runs two sequential LLM stages and assembles the result:

1. Persona — generate_persona(main_ccd, sentences) produces a persona (name, age, gender, occupation, family background, interpersonal relationships, physical condition, lifestyle, chief complaint).
2. Cognitive patterns — generate_cognitive_patterns(persona, main_ccd, dialogue_excerpt) produces the background, session-specific CCD, preferences, and client characteristics.
3. Assemble — combines the persona, the passed-in main_ccd, and the generated patterns into a CarsCharacter.

The character is returned; the generate CLI owns all I/O (loading seeds, saving output).

Usage

Provide seeds as a JSON list and run the CLI:

patienthub generate generator=cars input_path=data/seeds/cars.json

A single illustrative seed ships at data/seeds/cars.json so the generator is runnable out of the box. It is authored for PatientHub, not from the CARS paper or any released dataset. To reproduce the paper's diverse-profile generation, replace main_ccd with entries from a CBT-grounded CCD library and mesc_sentences with statements sampled from MESC (Chu et al., 2025).

Configuration

Parameter	Type	Default	Description
agent_name	string	cars	Generator identifier
prompt_path	string	data/prompts/generator/cars.yaml	Path to prompt file
model_type	string	"OPENAI"	Model provider key
model_name	string	"gpt-4o"	Model identifier
temperature	float	0.7	Sampling temperature
max_tokens	int	8192	Max response tokens
max_retries	int	3	API retry attempts

Seed Record Format

Seeds live in data/seeds/cars.json as a JSON list. Each record is validated against CarsSeed before generation — one character is produced per record:

[
  {
    "main_ccd": {
      "name": "Abe",
      "relevant_history": "Abe's father left when he was 11; his critical mother had unrealistic expectations. He recently lost his job and went through a divorce.",
      "core_beliefs": ["I am incompetent.", "I am a failure."],
      "intermediate_beliefs": [
        "If I try hard things I'll fail, so it's safer to avoid challenges.",
        "If I ask for help, people will see how incompetent I am."
      ],
      "coping_strategies": ["Avoids asking for help.", "Avoids challenges and difficult tasks."]
    },
    "mesc_sentences": [
      "What if I run out of money? I can't stop thinking about the bills.",
      "I should be able to do this on my own without asking anyone for help.",
      "I should have tried harder; I just keep failing at everything."
    ],
    "dialogue_excerpt": [
      {
        "role": "Therapist",
        "content": "Would it help to look at one small step together, maybe even asking someone for a hand?"
      },
      {
        "role": "Client",
        "content": "Ask for help? No. I should be able to handle this myself."
      }
    ]
  }
]

Field	Type	Description
main_ccd	object	A CBT-grounded main CCD (beliefs, history, coping strategies)
mesc_sentences	list	At least three sampled counseling statements (MESC-style)
dialogue_excerpt	list	Counselor–client turns, each { "role", "content" }

Output Format

{
  "name": "Abe",
  "persona": {
    "name": "Abe",
    "age": "42 years old",
    "gender": "Male",
    "occupation": "Recently unemployed (former manager).",
    "family_background": "...",
    "interpersonal_relationships": "...",
    "physical_condition": "...",
    "lifestyle": "...",
    "chief_complaint": "..."
  },
  "background": "The current counseling topic is treatment goal setting ...",
  "main_ccd": {
    "name": "Abe",
    "relevant_history": "...",
    "core_beliefs": ["I am incompetent.", "I am a failure."],
    "intermediate_beliefs": ["..."],
    "coping_strategies": ["..."]
  },
  "session_ccd": {
    "theme": "goal setting",
    "structured_cbt_strategy": "Goal list or agenda setting",
    "automatic_thought": {
      "situation": "...",
      "cognition": "...",
      "reaction": "..."
    },
    "intermediate_belief": {
      "attitude": "...",
      "rule": "...",
      "assumption": "..."
    },
    "resistance_triggers": ["..."]
  },
  "preferences": { "positive": ["..."], "negative": ["..."] },
  "client_characteristics": {
    "possible_responses_under_different_emotions": ["..."],
    "other_client_characteristics": ["..."]
  }
}

Note on the Emotion-Utterance Corpus

The paper references a "pre-defined emotion-utterance corpus" (§3.1.2) used at response time. The authors do not publish it, so the generator does not produce one and CARS characters do not carry emotion-utterance examples. See the CARS client for details.