XInsight: Integrative Stage-Consistent Psychological Counseling Support Agents for Digital Well-Being

This paper introduces XInsight, a multi-agent framework that aligns psychological support with the Exploration-Insight-Action paradigm through a structured Reason-Intervene-Reflect cycle to enhance interpretability and therapeutic effectiveness in digital well-being applications, accompanied by the XInsight-Bench evaluation protocol.

The Gist

Imagine you are feeling down, anxious, or stuck in a difficult situation. In the past, you might have gone to a therapist's office. Today, you might try talking to a chatbot on your phone. But here's the problem: most chatbots are like badly trained interns. They might be friendly, but they often forget what you said five minutes ago, they don't know how to help (do you need a pep talk, a logic check, or a breathing exercise?), and they don't leave you with a clear plan for tomorrow.

Enter XInsight. Think of XInsight not as a single robot, but as a highly organized, specialized medical team working together to help you, all inside your computer.

Here is how it works, broken down into simple parts:

1. The Problem with Current Chatbots

Most AI counselors today are like a one-trick pony. They try to solve every problem with the same conversation style. They don't follow the natural flow of healing, which usually happens in three distinct phases:

1. Getting to know you (Exploration).
2. Figuring out the root cause and trying new ways to think (Insight).
3. Making sure you stay healthy and don't fall back into old habits (Action).

Current bots often skip steps or mix them up, leaving you feeling unheard or confused.

2. The XInsight Solution: A "Dream Team" of Agents

XInsight is built like a specialized hospital ward where different experts handle different parts of your care. Instead of one AI trying to do everything, it uses a team of five "agents" (specialized AI programs) that pass your case file around like a relay race.

• The Detective (Exploration Agent):

  • Role: This agent is the warm, empathetic listener. Its job is to build trust, ask the right questions, and gather all the details about your life, history, and current worries.
  • Analogy: Imagine a detective who doesn't just listen to your story but organizes it into a neat, clear Case File so no detail is lost.

• The Triage Nurse (Routing Agent):

  • Role: Once the Detective has the facts, this agent looks at your "Case File" and decides: Which type of therapy will work best for this specific person?
  • Analogy: Think of a doctor who decides if you need a surgeon, a physical therapist, or a nutritionist. XInsight can switch between three major therapy styles:
    • SFBT: Focusing on solutions and what's working.
    • CBT: Challenging negative thoughts and logic.
    • MBCT: Using mindfulness and staying present.

• The Specialist (Therapeutic Agent):

  • Role: This is the expert who actually does the work. If the Triage Nurse said "CBT," this agent uses CBT techniques. If they said "Mindfulness," it switches gears.
  • Analogy: This is the specialist surgeon or therapist who performs the specific procedure needed, using the right tools for the job.

• The Coach (Consolidation Agent):

  • Role: After the hard work is done, this agent helps you wrap up. It asks, "What did we learn?" and "How do we keep this going?"
  • Analogy: Like a personal trainer helping you create a home workout plan so you don't lose your gains once you leave the gym. It helps you build a Relapse Prevention Plan (a safety net for when things get tough again).

• The Scribe (Recording Agent):

  • Role: This is the most unique part. While you talk, this agent is constantly turning your messy, emotional conversation into structured, professional medical notes.
  • Analogy: Imagine a court stenographer who doesn't just write down words, but instantly organizes them into a legal brief, a medical chart, and a to-do list. This ensures that if you come back next week, the system remembers exactly where you left off, not just "we talked about sadness."

3. The Three-Stage Journey

XInsight forces the conversation to follow a natural, healthy path:

• Stage 1: Exploration (The "Hello" Phase)
  • The team builds trust and creates a clear map of your problems. No jumping to solutions yet!
• Stage 2: Insight (The "Aha!" Phase)
  • The team picks the right therapy style (like CBT or Mindfulness) and helps you see your problems from a new angle. They challenge your negative thoughts or help you find your strengths.
• Stage 3: Action (The "Future" Phase)
  • The team helps you practice new skills and creates a plan for the future so you don't get stuck in the same loop again.

4. Why This Matters

The researchers didn't just build this; they tested it rigorously. They created a "gym" for AI counselors called XInsight-Bench to see if the bots actually help people.

• The Result: XInsight outperformed other AI systems. It had longer, deeper conversations, understood the different types of therapy better, and provided much clearer, more useful summaries of what happened.
• The Big Picture: XInsight isn't trying to replace human therapists. It's trying to be the best possible digital assistant for your mental well-being. It ensures that when you talk to a computer about your feelings, that computer acts like a trained professional who remembers your history, knows the right tools to use, and leaves you with a clear plan for tomorrow.

In short: XInsight turns a chaotic chat into a structured, professional, and helpful journey toward feeling better, just like a real therapy session, but available anytime, anywhere.

Technical Summary

Here is a detailed technical summary of the paper "XInsight: Integrative Stage-Consistent Psychological Counseling Support Agents for Digital Well-Being."

1. Problem Statement

Current Large Language Model (LLM) driven chatbots for psychological support suffer from three critical limitations when applied to web-based digital well-being:

• Lack of Paradigm Alignment: Most systems operate as single-stage, static interventions rather than modeling the dynamic, multi-stage nature of human counseling (Exploration, Insight, Action).
• Therapeutic Opacity: They often lack grounding in established therapeutic practices (e.g., SFBT, CBT, MBCT) and fail to adapt their strategy based on the client's evolving state.
• Unstructured Outputs: Interactions generate unstructured text, lacking the rigorous documentation (case formulations, relapse plans) required for continuity of care, interpretability, and accountability.

2. Methodology: The XInsight Framework

XInsight is a multi-agent framework that models psychological counseling as a stage-consistent workflow aligned with the classical Exploration-Insight-Action paradigm. It employs a unified Reason-Intervene-Reflect (RIR) cycle orchestrated by five specialized agents.

A. Agent Architecture

1. Exploration Agent ($A_{Explore}$):
   • Role: Builds trust, gathers information, and clarifies intentions.
   • Mechanism: Uses submodules like Mood Check, Psychoeducation, Goal Setting, Cognitive Coaching, and Behavioral Activation to construct a Case Conceptualization Form ($F_{case}$).
2. Routing Agent ($A_{Route}$):
   • Role: Performs Adaptive Therapeutic Routing (ATR).
   • Mechanism: Analyzes $F_{case}$ to infer dominant client patterns and selects the most suitable therapeutic school ($S^* \in \{SFBT, CBT, MBCT\}$) to guide the next stage.
3. Therapeutic Agent ($A_{Therapy}$):
   • Role: Executes school-specific interventions based on the selected $S^*$.
   • Mechanism:
     • SFBT: Uses techniques like Exception Finding, Scaling Questions, and Miracle Questions.
     • CBT: Focuses on Automatic Thoughts, Inter-belief Reorganization, and Core Belief Reconstruction.
     • MBCT: Cultivates Present Awareness, Acceptance, and Cognitive Defusion.
   • Outputs are recorded as a Therapeutic Record ($O_{ther}$).
4. Consolidation Agent ($A_{Consolidate}$):
   • Role: Integrates gains, reviews progress, and plans for the future.
   • Mechanism: Conducts Review Assessment, Skill Integration, and Relapse Prevention planning.
5. Recording Agent ($A_{Record}$):
   • Role: The "memory and documentation" layer.
   • Mechanism: Implements an Atomic Structured Tool-Taking scheme (inspired by Zettelkasten). It converts open-ended dialogue into standardized, structured psychological artifacts:
     • $F_{case}$ (Case Conceptualization)
     • $O_{ther}$ (Therapeutic Record)
     • $P_{rel}$ (Relapse Prevention Plan)
   • These artifacts are stored as dense embeddings with topic tags, enabling topic-guided, similarity-based retrieval for future sessions.

B. Workflow Dynamics

The system progresses through three distinct stages:

1. Stage 1 (Exploration): $A_{Explore}$ interacts with the client $\rightarrow$ $A_{Record}$ generates $F_{case}$.
2. Stage 2 (Insight): $A_{Route}$ selects a therapy school $\rightarrow$ $A_{Therapy}$ intervenes $\rightarrow$ $A_{Record}$ generates $O_{ther}$.
3. Stage 3 (Action): $A_{Consolidate}$ integrates $F_{case}$ and $O_{ther}$ $\rightarrow$ $A_{Record}$ generates $P_{rel}$.

3. Key Contributions

• Paradigm-Aligned Architecture: Recasts counseling into a three-stage, multi-agent workflow governed by a Reason-Intervene-Reflect cycle, ensuring stage-consistent control while maintaining dialogue fluidity.
• Adaptive Therapeutic Routing (ATR): Introduces a mechanism to dynamically switch between SFBT, CBT, and MBCT based on client needs, ensuring theoretical coherence.
• Structured Psychological Grounding: Transforms unstructured chat into interpretable clinical instruments ($F_{case}, O_{ther}, P_{rel}$), enabling longitudinal reasoning and accountability.
• XInsight-Bench & SGLE:
  • XInsight-Bench: A benchmark dataset with diverse client cases across demographics and therapeutic schools.
  • Scale-Guided LLM Evaluation (SGLE): A novel evaluation protocol combining therapy-specific clinical scales (FIT, CTS-R, MBCT-AS) with Human-Perspective Evaluation Criteria (HPEC) to assess quality objectively.

4. Experimental Results

The framework was evaluated on XInsight-Bench using various LLM backbones (e.g., Qwen3, Llama 3.1, Gemma 3) and compared against baselines like CBT-LLM and CACTUS.

• Performance Across Architectures: Qwen3 (14B) achieved the highest scores across all three therapeutic schools:
  • SFBT (FIT): 74.92 (vs. 69.94 for Qwen2.5).
  • CBT (CTS-R): 57.96 (vs. 54.53 for Qwen2.5).
  • MBCT (MBCT-AS): 22.17 (vs. 19.91 for Qwen2.5).
  • Insight: Paradigm-driven orchestration proved more critical than model size alone; smaller models struggled with cross-school generalization.
• Comparison with Baselines: XInsight significantly outperformed existing multi-agent systems (CBT-LLM, CACTUS) in both clinical scale scores and human-rated dimensions (Professionalism, Communication, Safety).
• ATR Accuracy: The routing mechanism achieved high accuracy (F1 > 0.89), with near-perfect distinction for MBCT (F1 = 0.98).
• Ablation Studies:
  • Removing Stage 2 (Insight) caused the most severe performance drop (-6.58 points), confirming its role as the critical bridge between exploration and action.
  • Replacing the Recording Agent with standard memory (MemGPT-style) resulted in a significant score drop (55.55 vs. 57.96), proving the necessity of structured psychological tool-taking.

5. Significance

• Digital Well-Being: XInsight provides a practical blueprint for integrating counseling-inspired agents into web applications, moving beyond "chit-chat" to structured, therapeutic support.
• Interpretability & Continuity: By generating standardized artifacts (Case Forms, Relapse Plans), the system bridges the gap between AI interaction and professional mental health documentation, facilitating human-in-the-loop oversight.
• Reproducible Evaluation: The introduction of XInsight-Bench and the SGLE protocol sets a new standard for evaluating AI in mental health, moving away from heuristic metrics toward clinically validated scales.
• Scalability: The framework demonstrates that multi-agent systems can effectively simulate complex, multi-school therapeutic workflows, offering a scalable solution for expanding access to mental health resources.