MIND: An Evidence-State Decision Interface for Criteria-Grounded Psychiatric Inquiry and Diagnosis¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/Lingxi-mental-health/MIND
Area: LLM Reasoning
Keywords: Psychiatric Consultation, Evidence-State Decision Interface, Criteria Grounding, Process Supervision, Reinforcement Learning
TL;DR¶
Addressing unsupported assertions and inquiry goal drift caused by unstructured text in multi-turn psychiatric consultation, MIND constructs a typed evidence state \(E_t\) encompassing observed evidence, unresolved criterion checks, active differentials, and reliability-gated clinical supports, sharing this interface across policy decisions, 3D clinical process rewards, information-gain scoring, and trajectory rectification to achieve up to 8.8-point diagnostic gains under EMR-grounded simulation.
Background & Motivation¶
Psychiatric consultation is one of the most intellectually demanding forms of multi-turn clinical dialogue. In real-world consultations, a physician must elicit discriminative clinical evidence from subjective, fragmented, and often metaphorical patient narratives, systematically evaluate competing diagnostic hypotheses against formal criteria such as DSM-5 or ICD-10, and determine when sufficient evidence has been accumulated to warrant committing to a diagnostic category. However, unlike standard symptom-checking dialogues, psychiatric complaints exhibit extensive syndromic overlap (e.g., high comorbidity between depression and anxiety), and decisive clinical signals—such as functional impairment severity, longitudinal course, bipolar/psychotic exclusion criteria, and acute safety risks—are easily overlooked under early signal ambiguity.
Existing conversational agents and retrieval-augmented generation (RAG) frameworks typically condition their policies directly on the raw dialogue history \(H_t\) or on unstructured retrieved reference passages appended to the prompt. This reliance on an implicit, text-entangled clinical state precipitates two recurring failure modes in clinical practice: first, models frequently make premature, unsupported assertions without systematically verifying whether necessary criterion thresholds or essential exclusion checks have been satisfied; second, agents suffer from severe inquiry drift in extended conversations, getting trapped in low-yield supportive loops or asking generic, redundant questions that fail to yield information gain on the active differential diagnoses.
The root cause of these limitations lies in the structural entanglement of four distinct clinical control variables within unstructured token sequences: observed patient facts, unresolved criterion checks, active differential hypotheses, and the reliability of external retrieved guidance. As the dialogue context expands, the policy is forced to repeatedly re-infer this multi-dimensional decomposition from scratch at every turn. To resolve this entanglement, this work discards the conventional view of retrieval as prompt context injection and instead models psychiatric consultation as an evidence-state Markov Decision Process (MDP). Core idea: establish a unified, typed clinical evidence-state decision interface (ESDI) \(E_t = (O_t, M_t, D_t, S_t, R_t)\), reframe knowledge retrieval as a reliability-gated state-construction operator, and share this operational state across policy reasoning, three-dimensional process rewards, information-gain estimation, and trajectory rectification to bound the decision space and enable criteria-grounded psychiatric inquiry.
Method¶
Overall Architecture¶
MIND operationalizes psychiatric consultation as a turn-budgeted task-oriented dialogue. In each interaction turn \(t\), the system avoids conditioning directly on the raw dialogue text \(H_t\). Instead, it executes an explicit two-step decision pipeline: first, the policy verbalizes dialogue facts into a clinical retrieval query \(q_t\), which is parsed into confirmed patient evidence \(O_t\), unresolved criterion and exclusion checks \(M_t\), and active differential hypotheses \(D_t\). Next, \(q_t\) queries an offline Psychiatric Reasoning Bank (PRB) constructed solely from training-split EMRs. Retrieved candidate entries are subjected to a strict reliability gate \(\text{Rel}(q_t, e_i) \ge \tau_{rel}\); admitted entries populate the criterion support slot \(S_t\) and reliability metadata slot \(R_t\), producing the full evidence state \(E_t = (O_t, M_t, D_t, S_t, R_t)\). Finally, the policy generates explicit chain-of-thought clinical reasoning conditioned on \((E_t, H_t)\) to produce the next inquiry or final diagnosis. During training, \(E_t\) simultaneously drives three-dimensional process rewards and information-gain optimization under GRPO, while at inference time, deterministic triggers monitor \(E_t\) to execute state-conditioned trajectory rectification.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Dialogue History Ht"] --> B["Stage 1: Typed Evidence-State Construction<br/>Extract qt and parse Ot, Mt, Dt"]
B --> C["Stage 2: PRB Retrieval & Reliability Gating<br/>Cosine lookup filtered by threshold, populating St, Rt"]
C --> D["Typed Decision Interface Et = (Ot, Mt, Dt, St, Rt)"]
D --> E["Stage 3: Explicit Clinical Reasoning & Action<br/>Two-step inference generating thought zt and action at"]
D --> F["Stage 4: 3D Process Supervision & InfoGain Reward<br/>Score symptoms/differentials/logic, credit |Mt| reduction"]
D --> G["Stage 5: State-Conditioned Trajectory Rectification<br/>Detect repetition/low-yield, trigger self-retry or PRB fallback"]
E --> H["Next Clinical Inquiry or Final Diagnosis"]
Key Designs¶
1. Criteria-Grounded PRB Construction & Reliability Gating: Eliminating Contextual Noise and Hallucinatory Knowledge Injection
Standard medical RAG pipelines inject long, unstructured guideline passages into the prompt, often introducing irrelevant distractors or hallucinated diagnostic premises. MIND constructs an offline Psychiatric Reasoning Bank (PRB) strictly from the training-split consultations and diagnostic manuals. Each entry is explicitly structured as a label-free typed tuple \(e_i = (q_i, c_i, o_i, g_i, x_i, u_i, \rho_i)\), representing the compact retrieval state, criterion/exclusion check, patient evidence anchor, missing check candidate, safety/exclusion cue, next-inquiry clinical rationale, and reliability scores/hard conflict flags. During online inquiry, the policy uses \(q_t\) to retrieve top-\(k\) entries via cosine similarity. Candidates are filtered through a reliability evaluator requiring \(\text{Rel}(q_t, e_i) \ge \tau_{rel}\) (calibrated at \(\tau_{rel}=0.2\)). Qualified entries populate \(S_t\) and \(R_t\), while failed entries fall back to null support. This ensures external clinical knowledge acts as a validated state-construction operator rather than an unverified text snippet, completely preventing test-set label leakage.
2. Evidence-State Decision Interface Sharing Across Operators: Resolving Representational Inconsistency Among Policy, Reward, and Recovery
Rather than treating structured output merely as a prompt display template, MIND establishes \(E_t = (O_t, M_t, D_t, S_t, R_t)\) as the invariant operational object shared across all decision components. The policy network \(\pi_\theta(E_t, H_t)\) consumes \(E_t\) to execute structured state-to-action reasoning; the process-reward judge \(R_{proc}(E_t, a_t)\) inspects reasoning consistency against the exact fields of \(E_t\); the information-gain module computes state transitions directly over the set difference \(\Delta |M_t| = |M_t| - |M_{t+1}|\); and the trajectory recovery operator \(\Gamma_{recover}(E_t, a_t)\) monitors field stagnation in \(R_t\) and \(M_t\). Because \(O_t, M_t, D_t\) follow a fixed schema and \(|S_t| \le k\) is bounded by retrieval depth, the size \(|E_t|\) remains approximately constant as dialogue horizon \(T\) grows, shrinking the effective hypothesis space and preventing context-length degradation.
3. Three-Dimensional Clinical Process Supervision & Information-Gain RL: Overcoming Sparse Terminal Rewards and Shortcut Learning
Relying exclusively on terminal correctness rewards encourages models to adopt superficial guessing shortcuts instead of methodical clinical inquiry. MIND integrates dense process rewards optimized via Group Relative Policy Optimization (GRPO). At each turn, an LLM judge evaluates the reasoning trace \(z_t^{(2)}\) across three clinical dimensions on a standardized 0–2 rubric:
$\(R_{proc}(E_t, a_t) = \text{Norm}(S_{sym} + S_{diff} + S_{dec})\)$
where \(S_{sym}\) evaluates symptom timeline, severity, and functional impairment analysis; \(S_{diff}\) rewards weighing active differential categories and identifying unaddressed exclusion criteria; and \(S_{dec}\) assesses whether the selected action is clinically justified to resolve the current evidence gap. Additionally, an intrinsic information-gain reward \(R_{info} = \gamma \cdot \mathbb{I}(|M_t| - |M_{t+1}| > 0)\) directly incentivizes the policy to formulate questions that close open checks in \(M_t\).
4. State-Conditioned Trajectory Rectification: Preventing Inquiry Drift, Semantic Repetition, and Consultative Deadlocks
During prolonged consultations, language models frequently exhibit inquiry drift, semantic loops, or malformed outputs. MIND implements a deterministic, state-conditioned rectification mechanism that tracks operational triggers including format violations, semantic repetition against prior questions, generic low-yield inquiries, low retrieval reliability (\(R_t\) indicating unresolved conflicts), and budget non-compliance. Upon detecting a trigger, the agent first executes a single constrained self-retry conditioned on the diagnosed failure mode. If the retry fails and the fallback cap is not reached (capped at 1 retry per turn, 2 fallbacks per episode, and \(\le 3\) total interventions), the policy replaces the action with a structured reference inquiry derived from the nearest reliable PRB entry. This deterministic rule avoids training complex value rerankers while guaranteeing strict clinical inquiry bounds.
Loss & Training¶
MIND is trained in two sequential phases: 1. SFT Warm-Start: Using 8 \(\times\) NVIDIA A100 GPUs, the Qwen3 backbone is fine-tuned with LoRA (\(r=64, \alpha=32\), learning rate \(1 \times 10^{-4}\), effective batch size 128 across 8 GPUs, 3 epochs) on 13,890 turn-level examples derived from 1,200 training EMRs to establish compliant two-step state verbalization and reasoning syntax. 2. GRPO Policy Optimization: The policy is optimized using GRPO to maximize expected return \(J(\theta) = \mathbb{E}_{\tau \sim \pi_\theta} \left[ \sum_{t=1}^T r_t + r_{term} \right]\), where the terminal diagnostic correctness reward carries a weight of 5.0, format compliance 0.1, clinical reasoning 0.01, and information gain 0.005. Optimization runs for 1,000–2,000 policy updates (Actor learning rate \(5 \times 10^{-6}\), KL coefficient \(\beta=0.02\), clipping ratio \(\epsilon \in [0.10, 0.18]\), 4 rollouts per prompt) using vLLM for accelerated trajectory generation.
Key Experimental Results¶
Main Results¶
Evaluation is conducted across two EMR-grounded patient simulators (PsySim-Std and PsySim-Adapt) built upon 400 held-out clinical records across four ICD-10 screening categories (Depression, Anxiety, Mixed Anxiety-Depression, Other), as well as on the independent public MDD-5k benchmark under a turn budget of \(L=10\).
Table 1: Per-class precision, recall, F1 (%), and overall diagnostic accuracy under patient simulators (Quoted from Paper Table 2)
| Setting | Method | Dep F1 | Anx F1 | Mix F1 | Oth F1 | Overall Acc (%) |
|---|---|---|---|---|---|---|
| PsySim-Std | GPT-4o (Inference-only) | 60.5 | 50.7 | 45.0 | 37.0 | 49.5 |
| PsySim-Std | DeepSeek-V3.2 (Inference-only) | 45.6 | 33.5 | 45.9 | 59.0 | 46.0 |
| PsySim-Std | Qwen3-8B (Zero-shot) | 62.5 | 40.2 | 31.0 | 31.2 | 43.0 |
| PsySim-Std | MRD-RAG (Retrieval-augmented) | 60.5 | 50.5 | 40.0 | 31.0 | 46.0 |
| PsySim-Std | Qwen3-8B† (SFT warm-start) | 56.7 | 57.1 | 49.9 | 50.3 | 53.4 |
| PsySim-Std | DoctorAgent-RL (Multi-agent RL) | 64.7 | 62.2 | 54.7 | 58.7 | 60.2 |
| PsySim-Std | DDO (Dual-decision RL) | 64.7 | 61.7 | 55.2 | 59.0 | 60.0 |
| PsySim-Std | Ours (MIND-4B) | 60.4 | 48.4 | 38.5 | 55.7 | 52.0 |
| PsySim-Std | Ours (MIND-8B) | 73.7 | 65.5 | 63.5 | 72.5 | 69.0 |
| PsySim-Adapt | GPT-4o (Inference-only) | 53.5 | 41.0 | 37.0 | 27.5 | 40.5 |
| PsySim-Adapt | DoctorAgent-RL | 58.2 | 53.2 | 42.5 | 52.4 | 52.5 |
| PsySim-Adapt | DDO | 54.2 | 50.5 | 46.0 | 48.5 | 49.4 |
| PsySim-Adapt | Ours (MIND-8B) | 68.0 | 52.0 | 46.5 | 70.5 | 59.8 |
Ablation Study¶
Table 2: Ablation of key components and state-representation controls (Quoted from Paper Table 5, Table 6, Table 10)
| Configuration / State Variant | PsySim-Std Acc (%) | PsySim-Std Macro-F1 (%) | PsySim-Adapt Acc (%) | MCR-shared ↑ (%) | Repetition ↓ (%) |
|---|---|---|---|---|---|
| MIND (Full System) | 69.0 | 68.8 | 59.8 | 54.2 | 8.0 |
| -- Component Ablations (Table 6) -- | |||||
| w/o SFT warm-start | 62.7 | 64.0 | 53.7 | - | - |
| w/o InfoGain reward | 66.5 | 67.7 | 57.5 | - | - |
| w/o Clinical reasoning reward | 65.3 | 66.9 | 56.0 | - | - |
| w/o PRB knowledge retrieval | 65.7 | 66.1 | 56.3 | - | - |
| w/o Reliability gate \(\tau_{rel}\) | 66.5 | 67.2 | 57.1 | - | - |
| w/o CoT thinking supervision | 61.5 | 60.6 | 51.9 | - | - |
| w/o Trajectory fallback | 67.1 | 67.8 | 58.0 | - | - |
| -- Matched State Controls (Table 5/10) -- | |||||
| Raw-history RL | 63.4 | 64.5 | 53.9 | 43.5 | 14.5 |
| Standard RAG RL | 66.0 | 65.8 | 56.5 | 48.3 | 12.2 |
| Structured-prompt RAG RL | 66.7 | 66.5 | 57.2 | 49.5 | 11.5 |
| One-sided: Policy=ES, Reward=Raw | 67.2 | 67.2 | 57.8 | 50.8 | 10.2 |
| One-sided: Policy=Raw, Reward=ES | 66.4 | 66.0 | 56.9 | 49.0 | 11.8 |
| Evidence-state RL (Both, no fallback) | 68.2 | 67.5 | 59.0 | 53.0 | 8.8 |
Key Findings¶
- Shared Evidence-State Representation Drives the Core Gain: In strictly controlled experiments holding the backbone, training data, RL steps, and retrieval content constant (Table 5/10), formatting the prompt into structured sections (Structured-prompt RAG RL) provides only a marginal boost (+0.7% Acc). In contrast, sharing \(E_t\) across both policy and reward computation yields a decisive jump to 68.2% (+1.5% Acc over prompt-only), proving that gains stem from consistent operational semantics across operators rather than prompt formatting or token length.
- Superior Missing-Check Resolution and Reduced Redundancy: MIND improves the Missing-Check Resolution rate under a shared clinical rubric (MCR-shared) to 49.6% (and 54.2% on MCR-self), substantially outperforming DDO's 41.2%, while slashing semantic question repetition from 18.4% to 8.0%.
- High Support Faithfulness and Physician Preference: Evaluated by DeepSeek-V3.2 and cross-validated by GPT-4o, MIND achieves the highest Support Faithfulness score (8.3 vs. DDO 7.5). In blind human evaluations by five medical experts across 160 cases, MIND secured a 78.5% pairwise win rate, compared to 52.5% for DDO and 45.0% for GPT-4o.
Highlights & Insights¶
- Reframing Retrieval as a State-Construction Operator: Rather than treating external literature as text appended to prompts, MIND decomposes clinical knowledge into typed schema records and maps them into explicit state variables via reliability thresholding, bridging the gap between open-domain RAG and formal POMDP decision making.
- Deterministic Trajectory Rectification Synergized with RL: By monitoring rule-based triggers (stagnation in \(|M_t|\), semantic duplicate loops, and reliability flags), MIND enforces self-retry and structured PRB fallback, providing a robust behavioral floor that prevents dialogic failure modes without requiring complex auxiliary reward models.
- Versatile Clinical Inquiry Schema: Factoring patient discourse into confirmed facts (\(O_t\)), unverified criterion checks (\(M_t\)), active differential hypotheses (\(D_t\)), external evidence (\(S_t\)), and confidence flags (\(R_t\)) provides a universal blueprint readily adaptable to other multi-symptom medical domains, such as neurology and rheumatology.
Limitations & Future Work¶
- Coarse Four-Way Screening Taxonomy: The framework is currently evaluated on a four-class screening setup (Depression, Anxiety, Mixed, Other) rather than granular clinical ICD-10/DSM-5 diagnostic sub-codes; the "Other" category encompasses heterogeneous conditions including insomnia and adjustment disorders.
- Potential Error Cascades in State Parsing: Although expert audits show a 91.8% partial-or-better accuracy for \(E_t\), active differentials (\(D_t\)) exhibit lower full accuracy (71.0%) during ambiguous initial turns, which could propagate downstream inquiry misdirection.
- Reliance on Simulated Patient Environments: The primary multi-turn interactions are evaluated using EMR-grounded dialogue simulators; prospective, double-blind clinical trials involving real patients and practicing clinicians are essential before clinical deployment.
Related Work & Insights¶
- vs DDO (Jia et al., 2025b): DDO employs dual-agent reinforcement learning for question selection and diagnosis, but operates over unstructured textual dialogue histories; MIND leverages a unified, typed evidence state, achieving superior criteria grounding and higher diagnostic accuracy (69.0% vs. 60.0%) with lower interactive complexity.
- vs MRD-RAG (Sun et al., 2025): MRD-RAG incorporates turn-level passage retrieval into dialogue prompts but lacks reliability gating and typed slot assignment; MIND demonstrates that reliability-gated state construction significantly reduces unsupported assertions and elevates support faithfulness from 6.3 to 8.3.
- vs POMDP Belief States & Dialogue State Tracking (Young et al., 2013; Wu et al., 2019): Traditional DST tracks static slot-value pairs or probability distributions over discrete categories; MIND's \(E_t\) dynamically maintains unresolved criteria checklists and external evidence anchors, providing a purpose-built state abstraction for complex diagnostic differential reasoning.
Rating¶
- Novelty: ⭐⭐⭐⭐⭐ Formulates an innovative evidence-state decision interface (ESDI) that recasts medical dialogue as an evidence-state MDP with state-shared policy, rewards, and recovery.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous evaluation across two patient simulators, external MDD-5k validation, matched representation controls, field-masking ablation, cross-family judge audits, and physician blind reviews.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally clear mathematical formulation, rigorous clinical definitions, and disciplined narrative presentation.
- Value: ⭐⭐⭐⭐⭐ Establishes a highly influential blueprint for grounding long-horizon conversational agents in formal diagnostic criteria and interpretable clinical reasoning.