Mindora: Clinical Mental-Health Reasoning Model

• Mindora is a clinically aligned Qwen3-based language model that uses hybrid SFT–RL training to generate structured and reliable mental-health reasoning.
• It operationalizes clinical reasoning by appraising narratives, diagnosing conditions, and planning interventions using specialized benchmarks and curated datasets.
• The model integrates benchmark design, trajectory generation, and consistency-aware reward mechanisms to achieve superior reasoning reliability across diverse mental-health tasks.

Mindora is the central model in the MentraSuite framework: an 8B-parameter, Qwen3-based LLM post-trained specifically for clinically aligned mental-health reasoning. Its intended role is not merely to generate supportive or knowledgeable text, but to perform step-wise, structured, and internally consistent reasoning over mental-health narratives drawn from social media, counseling dialogues, clinical-style vignettes, research abstracts, and misinformation-laden media transcripts. The model is optimized for appraisal, diagnosis-like judgments, intervention planning, abstraction of evidence, and verification of information, with explicit emphasis on reasoning reliability rather than final-answer accuracy alone (Xiao et al., 10 Dec 2025).

1. Position within MentraSuite

Mindora is one of three core elements in MentraSuite, alongside MentraBench and a Reasoning Trajectory Generation pipeline. In that framework, MentraBench provides the evaluation substrate; the trajectory-generation pipeline produces high-quality reasoning chains; and Mindora is the post-trained model that operationalizes those resources through hybrid supervised fine-tuning and reinforcement learning. The paper reports three model variants: Mindora_SFT, trained only by supervised fine-tuning on generated trajectories; Mindora_SFT+RL, trained by sequential SFT followed by RL; and Mindora_CHORD, the final model referred to as Mindora in the principal results (Xiao et al., 10 Dec 2025).

The model is explicitly framed as a reasoning-first psychological LLM. This positioning distinguishes it from systems oriented primarily toward empathy, emotional understanding, or knowledge recall. The governing objective is clinically aligned reasoning quality: concise, coherent, grounded, task-faithful, and internally consistent chains of thought that support appraisal, diagnosis, counseling-strategy selection, psychiatry question answering, evidence synthesis, and misinformation detection.

A plausible implication is that MentraSuite treats domain adaptation as a three-part problem rather than a single fine-tuning step: benchmark design, trajectory construction, and post-training are all coupled. That coupling is central to how Mindora is defined in the paper.

2. Benchmark structure and task taxonomy

MentraBench spans five core reasoning aspects, six tasks, and 13 datasets. It evaluates both task performance and reasoning quality, and it is designed to cover synthetic and real clinical-like text, social media, PubMed-derived materials, Cochrane-style evidence summaries, and video transcripts (Xiao et al., 10 Dec 2025).

Reasoning aspect	Task	Datasets
Appraisal	Cognitive error identification	CognitiveReframing; PatternReframe; Therapist Q&A
Diagnosis	Mental-health condition detection	DepSign; SWMH; T-SID
Intervention	Counseling strategy formulation	PsyDTCorpus$_M$; AnnoMI$_M$
Multi-step clinical reasoning	Psychiatry QA	MHQA; MedQA$_M$; MedMCQA$_M$; PubMedQA$_M$
Abstraction	Psychiatry Systematic Review Summarization	PSRS*
Verification	Mental-health misinformation identification	MentalMisinfo

The benchmark’s reasoning-quality evaluation is orthogonal to task correctness. Each reasoning trajectory is human-evaluated along five dimensions, each scored as binary: reasoning conciseness, logical coherence, hallucination avoidance, task understanding, and internal consistency. Their average defines a trajectory-level reliability score. This design reflects the paper’s claim that reliable mental-health reasoning requires not only correct outputs but also disciplined intermediate reasoning.

Several datasets have specialized construction or curation procedures. PsyDTCorpus$_M$ is used for training but excluded from evaluation. Psychiatry subsets are manually filtered from MedQA, MedMCQA, and PubMedQA to ensure mental-health relevance. PSRS* consists of manually curated Cochrane abstracts with expert scoring points. MentalMisinfo contains human-labeled video transcripts from YouTube and BitChute. The reported data sizes further indicate broad heterogeneity: for example, CognitiveReframing is listed as 751/173/284, PatternReframe as 1302/661/300, Therapist Q&A as 1662/409/507, MHQA as 1200/300/717, PSRS* as -/-/115, and MentalMisinfo as 123/31/130.

3. Model architecture and post-training objective

Mindora uses Qwen3-8B as the base policy model and Qwen3-32B as an auxiliary verifier for internal-consistency detection. Its post-training follows a CHORD-style hybrid SFT–RL framework in which supervised fine-tuning and on-policy reinforcement learning are dynamically blended rather than cleanly separated into fixed phases (Xiao et al., 10 Dec 2025).

The total loss is defined as

$$\mathcal{L}_{\text{total}}(\theta) = (1 - \mu(t)) \cdot \mathcal{L}_{\text{GRPO}}(\theta) + \mu(t) \cdot \mathcal{L}_{\text{SFT}-\phi}(\theta).$$

The global mixture weight $\mu(t)$ is scheduled with a warmup of 200 steps and a decay of 400 steps, using $\mu_{\text{peak}} = 0.5$ and $\mu_{\text{valley}} = 0.02$. This schedule interpolates between SFT-dominant and RL-dominant regimes.

The SFT component uses token-wise weighting,

$_M$0

so that expert tokens receive the highest weight when the current policy is uncertain about them. The paper characterizes this as a parabolic weighting that reduces both overemphasis on already-mastered tokens and overfitting to very unlikely tokens.

The RL component uses a GRPO (PPO-like) loss driven by a composite reward. The reward is

$_M$1

This reward structure is unusually strict. If format validity, length validity, or consistency fails, the reward is zeroed. FormatValid requires the exact schema > ...reasoning...</think><answer>...final answer...</answer>. LengthValid requires the reasoning trajectory length to satisfy $_M$2. Consistency is determined by the 32B auxiliary model and is defined to detect contradictions between earlier and later reasoning steps, conclusions unsupported by the stated evidence, and reasoning statements that conflict with domain knowledge given in the prompt. Quality is task-specific: exact correctness for single-choice tasks, Jaccard similarity for multi-label tasks, and scoring-point coverage for short-answer evidence abstraction.

The training setup reported for Mindora uses $_M$3 rollouts per prompt, temperature $_M$4, PPO clipping parameter $_M$5, and Adam with learning rate $_M$6, $_M$7, and $_M$8. Checkpoints are saved every 10 steps. A plausible implication is that the model’s alignment is enforced not by a scalar accuracy reward alone, but by a product of structural, epistemic, and task-level constraints.

4. Reasoning trajectory generation and rewriting

Mindora’s supervision is derived from a dedicated Reasoning Trajectory Generation pipeline rather than directly from raw task labels. The first stage is difficulty-based sample selection. For each dataset’s training split, Llama-3-8B-Instruct is run in zero-shot QA mode, and only the examples it answers incorrectly are retained. The retained cases form a hard pool intended to preserve training data with sufficient difficulty (Xiao et al., 10 Dec 2025).

For each hard case, GPT-4o generates an initial chain of thought and answer, and a second GPT-4o instance acting as a medical verifier checks whether the answer matches the ground truth. If it does not, the system refines the trajectory using one of four strategies: backtracking, exploring new paths, verification, or correction. The search is bounded by at most $_M$9 refinement iterations per attempt and at most $_M$0 attempts per sample; samples for which no correct solution is found are discarded.

Raw trajectories are then rewritten into a strict, consistency-oriented schema. The reasoning content must appear inside <think> ... </think>, organized by domain-specific subtitles such as ### Symptom Analysis, ### Differential Diagnosis, ### Risk Assessment, or ### Evidence from Study Results. The reasoning phase must end with ### Final Conclusion, which summarizes the logic and states the conclusion. The answer content appears separately inside <answer> ... </answer>, ending with Answer: [option/result]. The final answer must be logically consistent with the final conclusion.

This formatting serves several stated purposes. It removes messy backtracking and self-corrections from the final supervision signal, improves logical coherence by forcing modular structure, and explicitly links reasoning to the final answer to reduce mismatches between intermediate analysis and terminal output. The paper presents this rewriting as a mechanism for teaching clinically aligned decomposition, including symptom clustering, differential exclusion, risk evaluation, treatment selection, and evidence synthesis.

5. Empirical performance and reasoning reliability

Across 20 evaluated LLMs, Mindora_CHORD attains the highest overall average $_M$1 of 0.6933 on MentraBench. Its reported scores include $_M$2 for appraisal, $_M$3 for diagnosis, $_M$4 for psychiatry QA, 0.8379 on PSRS, and 0.7178 on MentalMisinfo. On psychiatry QA, the paper reports that Mindora_CHORD exceeds GPT-o4-mini (0.7402) and DeepSeek-R1 (0.7459). It also improves substantially over its own training ablations and backbone: Qwen3-8B = 0.5729, Mindora_SFT = 0.6367, and Mindora_SFT+RL = 0.6548 (Xiao et al., 10 Dec 2025).

The paper also reports a dedicated trajectory-quality comparison. Averaged over the five human-evaluated dimensions, Mindora_CHORD reaches 0.9731. The corresponding scores listed in the paper include DeepSeek-R1 = 0.9827, GPT-o4-mini = 0.8731, Qwen-plus = 0.9270, Psyche-R1 = 0.9442, and Qwen3-8B = 0.9039. The interpretation given is that Mindora_CHORD especially improves conciseness and coherence relative to the base model, which is consistent with the structured-trajectory and consistency-aware reward design.

These results matter because they separate two questions often conflated in mental-health LLM evaluation: whether a model arrives at the right answer, and whether it reasons in a clinically usable way. Mindora is presented as improving both. The paper does not claim uniform dominance across every submetric and every comparator, but it does claim the strongest average performance across the benchmark and a marked gain over the underlying 8B backbone.

6. Supported reasoning modes and clinical-style decomposition

Mindora is trained and evaluated to support several distinct forms of mental-health reasoning. In appraisal, it identifies cognitive distortions such as All-or-Nothing Thinking, Overgeneralization, Labeling, Fortune Telling, Mind Reading, Should Statements, Personalization, Mental Filtering, Catastrophizing, Emotional Reasoning, and Discounting the Positive (Xiao et al., 10 Dec 2025). The case study highlighted in the paper is diagnostically instructive: when a client says “Am I insane?” after feeling watched, many models classify the distortion as Emotional Reasoning, whereas Mindora identifies Labeling by focusing on the thought content rather than the emotion alone.

In diagnosis-like reasoning, the model is evaluated on social-media posts and exam-style vignettes requiring discrimination among depression, anxiety, suicidal ideation, and other mental disorders. The reported reasoning schema includes symptom extraction, duration, functional impact, differential diagnosis, risk assessment, and justified final classification. The paper describes this as mapping to DSM-like or ICD-like categories (implicitly) rather than as explicit rule-based diagnosis.

In intervention planning, Mindora selects among 13 counseling strategies: Clarification, Paraphrasing, Reflection of Feeling, Summarizing, Questioning, Immediacy, Silence, Self-Disclosure, Confrontation, Encouragement, Repetition, Interpretation, and Guidance. The task is not open-ended role-play; it is structured strategy identification conditioned on the client state, readiness to change, and risk indicators.

In multi-step clinical reasoning, the model operates over MHQA and psychiatry subsets of MedQA, MedMCQA, and PubMedQA. In abstraction, it performs Psychiatry Systematic Review Summarization by extracting PICO-style evidence structure and summarizing what the evidence shows. In verification, it detects misleading or non-evidence-based claims in noisy video transcripts. Across these settings, the same structured reasoning format is retained, typically with domain-specific subtitles followed by ### Final Conclusion and a short answer.

A plausible implication is that Mindora’s unifying contribution lies less in a single downstream task than in enforcing one reasoning protocol across heterogeneous mental-health problems. The paper treats that uniformity as part of the model’s reliability profile.

7. Limitations, safety, and disambiguation

The paper explicitly treats Mindora as a research model for reasoning and assessment rather than as a deployable clinical authority. It states that Mindora is not a licensed clinician and not a substitute for professional care (Xiao et al., 10 Dec 2025). The risks identified include misreading self-reports, exaggerating or downplaying symptoms, amplifying anxiety by treating subjective statements as facts, and producing inaccurate or misleading feedback in high-risk contexts such as suicidal ideation or psychosis. More general concerns include non-maleficence, therapeutic alliance, evidence-based practice, and regulatory gaps for wellness or therapy apps.

Mindora’s safety-oriented mitigations are architectural rather than dispositive. The reward’s Consistency(a) term penalizes factual and logical contradictions; the enforced <think>...<answer>...</answer> format and length bounds constrain uncontrolled verbosity; hallucination avoidance is directly evaluated; and task understanding is scored separately to discourage inappropriate shifts from classification or reasoning into unsolicited counseling behavior. Even so, the paper acknowledges residual failure modes on subtle or ambiguous cases, possible bias inherited from social-media corpora, and the lack of nonverbal context.

A recurrent source of confusion is nomenclature. Mindora should not be conflated with MINDSTORES, a different system described as “Memory-Informed Neural Decision Synthesis for Task-Oriented Reinforcement in Embodied Systems,” which concerns persistent natural-language experience memories for embodied agents in MineDojo/Minecraft rather than mental-health reasoning (Chari et al., 31 Jan 2025). The similarity in naming does not indicate a shared task domain, architecture, or evaluation protocol.

Taken together, Mindora is best understood as a post-trained, Qwen3-8B-based reasoning model whose distinctiveness derives from the integration of MentraBench, structured trajectory generation, and a hybrid SFT–RL objective with explicit inconsistency detection. Its significance lies in showing that, within mental-health NLP, trajectory quality, structural faithfulness, and internal consistency can be elevated to first-class training targets rather than treated as incidental byproducts of answer accuracy.