MIND: Unified Inquiry and Diagnosis RL with Criteria Grounded Clinical Supports for Psychiatric Consultation

The paper proposes MIND, a unified inquiry-diagnosis reinforcement learning framework that leverages a Criteria-Grounded Psychiatric Reasoning Bank and rubric-based process rewards to enhance diagnostic accuracy, mitigate inquiry drift, and ensure clinically supported reasoning in psychiatric consultations.

The Gist

Imagine you are trying to solve a very complex mystery, like a detective trying to figure out why a friend is acting strangely. In the world of mental health, this "detective work" is called a psychiatric consultation. The goal is to ask the right questions to understand the person's feelings, figure out what's wrong, and offer help.

The paper introduces a new AI system called MIND (Unified Inquiry and Diagnosis RL with Criteria Grounded Clinical Supports). Think of MIND as a super-smart, highly trained detective assistant designed specifically to help doctors (or AI doctors) solve these mental health mysteries without getting confused or making mistakes.

Here is how MIND works, explained through simple analogies:

1. The Problem: The "Lost Detective"

Current AI chatbots are great at chatting, but when it comes to diagnosing mental health issues, they often act like amateur detectives.

• The "Guessing Game": If a patient says, "I'm tired and sad," a normal AI might immediately jump to "You have depression!" without checking if they've been tired for two weeks or if they have other symptoms. It's like guessing the answer to a math problem without showing your work.
• The "Wandering Conversation": Sometimes, the AI gets distracted. The patient talks about their boss being mean, and the AI starts asking about their boss instead of digging deeper into the medical symptoms. The conversation drifts off-topic, and the AI misses the crucial clues needed for a diagnosis.

2. The Solution: The "MIND" System

MIND is built to fix these two problems. It acts like a detective who has a rulebook, a library of past cases, and a GPS to keep them on track.

A. The "Rulebook Library" (The Psychiatric Reasoning Bank)

Imagine a detective has a massive library of past cases and a strict rulebook (like the official medical guidelines).

• How MIND uses it: Before asking a question, MIND doesn't just guess. It looks up its "library" (called the Psychiatric Reasoning Bank or PRB).
• The Analogy: If a patient says, "I haven't slept well," MIND checks its rulebook. The rulebook says, "To diagnose depression, we need to know if this has lasted more than two weeks." So, MIND asks, "How long has this been happening?" instead of just saying, "That sounds sad."
• The Result: MIND never makes a diagnosis without checking the specific rules first. It prevents "unsupported guesses."

B. The "GPS for Questions" (Stopping the Drift)

Imagine you are driving to a destination, but the GPS keeps telling you to turn left into a scenic park that has nothing to do with your destination.

• The Problem: Normal AIs often get stuck in "empathy loops." They say, "That sounds really hard," and then ask, "Tell me more," but they don't ask specific questions to get medical facts. They drive in circles.
• How MIND uses it: MIND has a GPS (called Value-Aware Trajectory Rectification). It constantly checks: "Is this question helping me get closer to the diagnosis?"
• The Analogy: If the AI starts asking about the patient's favorite movie (which is nice but not medically useful), the GPS says, "No, that's a detour! Let's go back to the main road." It gently steers the conversation back to the symptoms that matter, ensuring every question brings them closer to the answer.

C. The "Coach" (Process Rewards)

Imagine a student taking a test. A normal teacher only grades the final answer (Right/Wrong). But MIND has a coach who watches every step of the student's thinking.

• How it works: MIND is trained using a method where it gets points not just for the final diagnosis, but for how it got there. Did it check the right symptoms? Did it rule out other possibilities?
• The Analogy: It's like a chess coach who says, "Good move, but you missed a checkmate opportunity three turns ago." This helps MIND learn to think more logically and carefully, rather than just guessing the final result.

3. Why is this a Big Deal?

Mental health is tricky because people describe their feelings in messy, confusing ways.

• Old AI: "You seem sad. Maybe it's depression." (Risky, might be wrong).
• MIND: "You seem sad. Let's check: Is this sadness lasting more than two weeks? Does it stop you from working? Have you had any thoughts of hurting yourself? Okay, based on these specific rules, it looks like Depression, but we need to rule out Anxiety first."

Summary

MIND is like upgrading a chatbot from a friendly conversationalist to a rigorous medical detective.

1. It uses a digital library to ensure every question follows medical rules.
2. It uses a GPS to stop the conversation from wandering off-topic.
3. It uses a coach to reward logical thinking, not just the final answer.

The result is an AI that is safer, more accurate, and better at helping doctors understand what a patient is really going through, without making dangerous guesses.

Technical Summary

Here is a detailed technical summary of the paper "MIND: Unified Inquiry and Diagnosis RL with Criteria Grounded Clinical Supports for Psychiatric Consultation."

1. Problem Definition

The paper addresses the challenges of using Large Language Models (LLMs) for psychiatric consultation, a task significantly more complex than general medical diagnosis due to:

• Subjective Ambiguity: Patient narratives are often metaphorical, inconsistent, or underspecified.
• Comorbidity Complexity: Symptoms often overlap (e.g., depression and anxiety), requiring rigorous differential reasoning.
• Two Fundamental Challenges in Existing Methods:
  1. Weak Criteria Grounding: Models often make unsupported clinical assertions when symptoms are atypical, ignoring diagnostic thresholds (e.g., duration, functional impairment) and exclusion rules.
  2. Inquiry Drift: In multi-turn dialogues, models struggle to maintain focus, often engaging in low-yield questioning, off-topic empathy loops, or failing to gather critical diagnostic cues efficiently.

2. Methodology: The MIND Framework

The authors propose MIND, a unified Inquiry–Diagnosis Reinforcement Learning (RL) framework. It integrates retrieval-augmented generation (RAG) with process-supervised RL to optimize multi-turn clinical dialogues.

A. Criteria-Grounded Psychiatric Reasoning Bank (PRB)

To mitigate hallucinations and unsupported assertions, MIND constructs a specialized knowledge base:

• Construction: Historical consultation cases are distilled into Clinical Retrieval States (fact-only, keyword-dense summaries) using an LLM (Kimi-K2).
• Content: Each entry pairs a retrieval state with a Clinical Support ($r_i$) derived from guidelines/textbooks. These supports explicitly specify:
  • Known clinical facts.
  • Critical information gaps or exclusion checks (e.g., "Is functional impairment present?").
  • Diagnostic rationales for the next inquiry step.
• Retrieval: At each turn, the agent converts the dialogue history into a retrieval state, fetches top-$k$ similar supports from the PRB, and uses them as grounded hints for the next question.

B. Explicit Clinical Reasoning with Process Supervision

Instead of generating a response directly, the agent follows a two-stage turn format:

1. Retrieval Query Generation: The agent outputs a structured query to fetch PRB supports.
2. Reasoning-then-Response: Conditioned on the retrieved supports, the agent generates an explicit reasoning trace (Chain of Thought) before the final response.
   • The trace must cover: Symptom analysis, Differential considerations, and Decision logic.
   • Rubric-Based Process Rewards: An LLM judge scores the reasoning trace on three dimensions: Symptom Coverage ($S_{sym}$), Differential Focus ($S_{diff}$), and Decision Logic ($S_{dec}$). This provides dense, fine-grained supervision for intermediate steps, not just the final diagnosis.

C. Value-Aware Trajectory Rectification

To prevent inquiry drift and low-utility loops in noisy dialogues:

• Mechanism: The system monitors turn-level utility. If a turn is detected as low-yield (e.g., repetition, format failure, off-topic), it triggers interventions.
• Interventions:
  • Self-Retry: The agent regenerates the turn under stricter constraints.
  • PRB-Guided Fallback: If self-retry fails, the system forces a return to a structured path by injecting a reference inquiry from the PRB (e.g., standard SCID-5 style questions for severity or risk).

D. Policy Optimization

MIND is trained via a staged pipeline:

1. Supervised Fine-Tuning (SFT): Warm-starts the policy using clinician-like trajectories and PRB artifacts to establish stable formatting and basic behavior.
2. Reinforcement Learning (RL): Uses GRPO (Group Relative Policy Optimization) with a mixed reward function:
   • Process Rewards: Based on the rubric scores of the reasoning trace.
   • Information Gain: Rewards acquiring new, decisive clinical cues.
   • Terminal Reward: Accuracy of the final diagnosis.
   • Penalties: For formatting errors, loops, or budget violations.

3. Key Contributions

1. Unified Framework: Introduced MIND, the first framework to jointly optimize inquiry policies and diagnostic decisions in psychiatry using retrieval augmentation and mixed RL rewards.
2. Psychiatric Reasoning Bank (PRB): Proposed a novel mechanism to convert unstructured guidelines into structured, criteria-grounded supports that anchor the agent's reasoning, reducing hallucinations.
3. Process-Supervised RL: Designed a training paradigm that couples explicit reasoning traces with rubric-based process rewards and value-aware trajectory rectification, ensuring stable multi-turn interactions.
4. Comprehensive Evaluation: Demonstrated superior performance in diagnostic accuracy, interpretability, and empathetic interaction quality compared to strong baselines.

4. Experimental Results

The model was evaluated on a dataset of 1,000 de-identified Electronic Medical Records (EMRs) covering Depression, Anxiety, Mixed Anxiety-Depression, and Other disorders.

• Diagnostic Accuracy: MIND (8B variant) achieved 71.5% Accuracy and 72.5% Macro-F1 on the standard simulator, significantly outperforming inference-only models (e.g., GPT-4o at 49.5%) and other RL-based baselines (e.g., DDO at 53.0%).
• Generalization: MIND maintained robust performance under a more challenging "Adapted" simulator (PsySim-Adapt) that introduces narrative variations, achieving 62.5% F1 vs. 45.7% for the next best baseline (DDO).
• Support Faithfulness: MIND scored highest (8.6/10) in "Support Faithfulness," meaning its clinical claims were most consistent with the retrieved guidelines and patient context, minimizing unsupported assertions.
• Human Evaluation: In a blind study with medical experts, MIND was preferred in 82% of comparisons over baselines, scoring highest in Professional Reliability (4.10/5) and Naturalness (3.95/5).
• Ablation Studies: Removing the PRB, reasoning supervision, or fallback mechanisms led to significant performance drops (e.g., removing reasoning supervision dropped F1 by ~12%), confirming the necessity of each component.

5. Significance

• Clinical Safety: By grounding inquiries in diagnostic criteria and enforcing exclusion checks (e.g., ruling out bipolarity before diagnosing depression), MIND reduces the risk of dangerous misdiagnoses common in "black box" LLMs.
• Efficiency: The trajectory rectification mechanism ensures that multi-turn dialogues remain high-yield, avoiding the "chitchat" or repetitive loops that plague current medical chatbots.
• Interpretability: The explicit reasoning traces provide a transparent audit trail for why a specific question was asked or a diagnosis was made, which is crucial for clinical adoption.
• Scalability: The framework demonstrates that combining structured knowledge retrieval with process-supervised RL can effectively handle the high ambiguity and complexity of psychiatric care, offering a blueprint for other complex medical domains.

In summary, MIND represents a significant step forward in AI-assisted psychiatry by moving beyond simple conversational fluency to criteria-grounded, diagnostically rigorous, and self-correcting clinical reasoning.