Beyond Empathy: Integrating Diagnostic and Therapeutic Reasoning with Large Language Models for Mental Health Counseling

1 Introduction

• Current counseling LLMs often lack explicit DSM/ICD-based diagnostic reasoning and rarely incorporate or adapt multiple therapeutic perspectives.These gaps limit reliability in clinical decision support or realistic therapy simulation; different patients may need different modalities.Source: §1 · pp. 1–2
• Its automated synthesis pipeline parses real-world mental health posts, generates multi-turn dialogues, simulates clinical reasoning, and filters data for quality.The filtering checks for incomplete thinking, context incoherence, mismatched responses, and missing frameworks.Source: §1 · p. 2
• PsyLLM is proposed as the first LLM to systematically integrate diagnostic and therapeutic reasoning for mental health counseling.It applies diagnostic criteria and psychotherapeutic frameworks during counseling interactions.Source: §1 · p. 2
• The reasoning is guided by international diagnostic standards such as DSM/ICD and diverse therapeutic strategies including CBT, ACT, and psychodynamic approaches.The model assesses emotions, analyzes cognitive patterns, and formulates strategies grounded in these criteria and modalities.Source: §1 · pp. 1–2
• The paper introduces a benchmark and evaluation protocol for counseling quality, alongside analysis of model performance and data characteristics.The benchmark assesses Empathy & Insight, Support & Autonomy, Attunement & Presence, and Safety & Boundaries.Source: §1 · p. 2

2.1 Mental Support Chatbots

• Mental-health chatbots progressed from empathy-oriented replies to multi-turn dialogue, strategy-guided role-play, and therapeutic annotations, but often lacked underlying clinical reasoning.Examples include SMILE (Qiu et al., 2024a), SoulChat (Chen et al., 2023), SweetieChat (Ye et al., 2025), and CPsyCoun (Zhang et al., 2024a).Source: §2.1 · pp. 2–3

2.2 Deep Reasoning in LLMs

• Reasoning research extends chain-of-thought methods into dialogue, emotional inference, and psychological tools, including questionnaires and therapist–AI collaboration.The cited work includes CoT prompting (Wei et al., 2022), agent dialogue (Shu et al., 2025), and psychological reasoning tools (Yang et al., 2023; Chen et al., 2025).Source: §2.2 · p. 3

3.1 Data Collection

• The dataset draws on pre-filtered Reddit corpora and patient utterances from two counseling datasets based on real-world cases and patient-doctor dialogues.The Reddit sources are Identifying-depression, Dreaddit, and LRF; the counseling sources are ChatCounselor and CPsyCoun.Source: §3.1 · p. 3

3.2 Data Parsing and Interaction Planning

• For each Reddit post, an LLM assesses emotions, sets one to three dialogue rounds based on issue complexity, and defines a therapeutic theme for each round.Themes guide round goals so the dialogue builds on earlier turns.Source: §3.2 · p. 3

3.3 Question Extraction from Dialogue

• Patient-utterance prompts come from simulated exchanges following Reddit posts and from patient turns extracted from existing counseling dialogues.For Reddit posts, the LLM first generates an empathetic counselor response and then a plausible patient reaction to use as a prompt.Source: §3.3 · p. 4

3.4 Response Generation with Diagnostic Standard & Therapeutic Strategy

• Given patient utterances and dialogue context, the LLM jointly generates a reasoning trace and counselor response guided by diagnostic standards and diverse therapeutic strategies.The response is conditioned on the generated reasoning, linking the clinical assessment and therapeutic rationale to the intervention.Source: §3.4 · pp. 4–5

3.5 Systematical Dialogue Validation

• Generated dialogue turns are retained only when they pass checks for reasoning completeness, context coherence, response alignment, and adherence to diagnostic and therapeutic guidance.The checks assess the reasoning trace and response against the patient utterance and preceding context.Source: §3.5 · p. 5

3.6 Dataset Statistics and Analysis

• The dataset contains 19,302 dialogues and 49,374 utterances, with patient utterances averaging 222 tokens and counselor responses averaging 519 tokens.Its distributions include multiple psychotherapy approaches and scene categories; moderate and severe cases are the most frequent severity levels.Source: §3.6 · pp. 5–6
• The Table 3 caption reports that OpenR1-Psy achieves the highest performance in a synthetic-data comparison across four metrics and normalized averages.The caption specifies metric maximum scores of 2, 4, 3, and 1.Source: Table 3 · p. 6

3.7 PsyLLM Training

• PsyLLM is trained with supervised fine-tuning on OpenR1-Psy, generating a validated reasoning trace before its corresponding counselor response.The training objective maximizes the likelihood of the concatenated reasoning-and-response target given the dialogue context and patient utterance.Source: §3.7 · p. 6

4.1 Experimental Settings

• PsyLLM was fine-tuned from Qwen3-8B, and counseling responses were evaluated on empathy, support, attunement, and safety dimensions using automatic metrics and expert review.Training used three epochs; evaluation assessed single turns containing both a reasoning trace and counselor utterance.Source: §4.1 · pp. 6–7

4.2 Synthetic Data Comparison

• OpenR1-Psy’s normalized average exceeded ChatCounselor by 57% and CPsyCounD by 41%, with gains across the four reported counseling dimensions.Against ChatCounselor, gains were 120% in Emp&In, 96% in Sup&Aut, 86% in Att&Pre, and 1% in Saf&Bou; against CPsyCounD, they were 71%, 84%, 44%, and 2%, respectively.Source: §4.2 · p. 7

4.3 Performance Comparison

• PsyLLM attained the highest automatic Normalized Average (0.607), ahead of GPT-4o (0.552) and DeepSeek-V3 (0.540), and led human ratings across all dimensions with a Normalized Average of 0.601.The automatic scores were highest for Emp&In (0.904), Att&Pre (1.440), and Saf&Bou (0.945); GPT-4o scored slightly higher than PsyLLM on Sup&Aut, 2.218 versus 2.210.Source: §4.3 · p. 7, Table 4 · p. 8
• PsyLLM significantly outperformed In-Context Learning and Two-Phase Prompting, supporting joint generation of reasoning and response over separate-stage prompting.The authors describe supervised fine-tuning as essential for embedding therapeutic reasoning and joint generation as coupling the clinical rationale to the intervention.Source: Figure 5 · p. 8, §4.3 · p. 8

4.4 Analysis

• Using the full OpenR1-Psy data produced the best ablation score (0.778), compared with 0.697 without therapeutic guidance and 0.740 without diagnostic guidance.The metric correlation analysis found Support & Autonomy correlated with Attunement & Presence at 0.73, while Safety & Boundaries correlations with the other dimensions ranged from 0.20 to 0.33.Source: §4.4 · pp. 8–9, Figure 6 · p. 9
• Qwen3-8B scored 0.757, above Qwen3-3B (0.738) and Qwen3-1.7B (0.675); with Qwen3-8B fixed, 80% of training data scored 0.746 versus 0.656 at 100%.The 20% data variant scored 0.718, and the Safety metric remained around 0.99 across data scales.Source: Figure 7 · p. 9, §4.4 · p. 9
• Psychology graduate students and Gemini rated PsyLLM’s reasoning traces highly, with normalized averages of 0.962 and 0.935, respectively.Humans scored higher overall and on most dimensions, while automatic evaluation scored higher on Coherence.Source: §4.4 · pp. 9–10

4.5 Case study

• In the case study, PsyLLM used DSM/ICD-informed reasoning to identify perfectionism and hypervigilance, then combined CBT reframing with humanistic self-compassion.The comparison describes CPsyCounX as offering generic coping suggestions without addressing the client’s underlying cognitive and emotional patterns.Source: Figure 8 · p. 10, §4.5 · p. 10

• PsyLLM is presented as addressing gaps in explicit DSM/ICD-aligned diagnosis and integration of diverse therapeutic modalities in mental health LLMs.Source: §5 · p. 10
• The OpenR1-Psy dataset is generated through an automated pipeline and includes reasoning traces guided by diagnostic standards and therapeutic frameworks such as CBT, ACT, and psychodynamic approaches.The paper says the generated dialogue data undergoes multidimensional filtering.Source: §5 · p. 10
• The authors report that PsyLLM significantly outperforms state-of-the-art baselines on their counseling-quality benchmark.Source: §5 · p. 10

• Early efforts emphasized single-turn empathy-oriented responses, including ChatCounselor and PsyQA, which generated supportive answers without modeling multi-turn therapeutic dynamics.Source: §A.1 · p. 14
• Later systems extended support to multi-turn contexts: SMILE expanded supportive utterances into conversations, SoulChat fine-tuned multi-turn empathetic dialogues, and SweetieChat used strategy-guided role-play.Source: §A.1 · p. 14
• Other approaches modeled reasoning or clinical tasks, including CPsyCoun’s therapeutic-type annotations, ESCoT’s emotion-and-strategy chain of thought, MentaLLaMA’s condition prediction with rationales, and PsyGUARD’s suicide-risk assessment.The review notes that CPsyCoun did not model the underlying reasoning behind its therapeutic strategies.Source: §A.1 · p. 14

A.2 Deep Reasoning in LLMs

• Prior work extends chain-of-thought reasoning beyond symbolic tasks to domains including medical, multimodal, emotional, and interactive dialogue reasoning.The review cites approaches using reinforcement learning, emotional perception, and psychological structures for dialogue or personality-related tasks.Source: §A.2 · p. 14
• The example trace identifies rejection, emotional swings, and feeling unheard, then plans validation and gentle exploration before producing an empathetic response.Source: §A.2 · p. 15

B Extended Multi-Turn Dialogue Case Study

• The extended case presents each counselor utterance after an explicit reasoning trace, showing how reasoning and responses evolve while maintaining context and adapting therapeutic strategy.The figure caption describes the trace as outlining a therapeutic assessment and strategy for each turn.Source: B Extended Multi-Turn Dialogue Case… · p. 15, Figure 9 · p. 15

Human evaluator

• Five psychology experts conducted the human evaluation.Compensation was based on the type and volume of work, in line with local regulations and professional standards.Source: C Details for Human evaluator · p. 15

Correlation Between Expert and Automatic Evaluations

• All four expert–Gemini-2.0-Flash score correlations were statistically significant across the counseling dimensions.The comparison used 500 randomly sampled test cases, evenly distributed across the four dimensions.Source: C Details for Human evaluator · p. 15

D Detailed Automatic Evaluation Results

• Table 7 presents complete automatic-evaluation results for PsyLLM and baseline models across topics.The supplied text identifies the table’s scope but gives no topic-level scores or comparisons.Source: D Detailed Automatic Evaluation Results · p. 15

E Ethics Statements

• The study’s first data portion comprises anonymized public posts, while the existing dialogue data were filtered, rewritten, and proofread to remove sensitive content.The dataset is described as using publicly available data; the dialogue-data processing included rule-based filtering and human review.Source: E Ethics Statements · p. 16
• PsyLLM was not supervised or validated by licensed clinicians, and fine-tuning lacked human feedback, leaving potential emotional-harm risks in sensitive scenarios.The authors state that responses should not replace professional mental health services.Source: E Ethics Statements · p. 16
• OpenR1-Psy is described as freely available for non-commercial academic research through an open-access repository.The statement says the dataset would be publicly released upon paper acceptance.Source: E Ethics Statements · p. 16

F Prompts for OpenR1-Psy

• OpenR1-Psy uses structured prompts for therapeutic interaction planning, empathic dialogue reconstruction, DSM/ICD-grounded therapeutic empathy, dialogue validation, and diagnostic classification.These prompt types are shown in Figures 10–14.Source: F Prompts for OpenR1-Psy · p. 16

G Automatic Evaluation

• Automatic evaluation uses several prompts covering dialogue quality and reasoning evaluation.Figures 15–17 show prompts for data evaluation, performance evaluation, and internal-thinking evaluation.Source: G Automatic Evaluation · p. 16
• The evaluation standard refers to the counseling-quality metrics listed in Table 12.Source: G Automatic Evaluation · p. 16
• The automatic-evaluation prompts include separate assessments for data, performance, and internal thinking.Source: G Automatic Evaluation · p. 16

H Model Comparison Examples

• Qualitative case studies compare conversational behavior across psychological dialogue models.Figures 18–23 show client dialogues with DeepSeek-V3, GPT-4o, CPsyCounX, ChatCounselor, MeChat, and PsyDTLLM.Source: H Model Comparison Examples · p. 16
• Table 7 gives full automatic-evaluation results for PsyLLM and baseline models; bold and underlined values mark the best and second-best values.The caption does not provide the underlying values or identify a winner in the supplied text.Source: Table 7 · p. 17
• Performance comparisons cover Qwen3-1.7B, Qwen3-3B, and Qwen3-8B base models.Source: Table 8 · p. 18
• A separate comparison examines performance under different data scales.Source: Table 9 · p. 18
• The ablations examine diagnostic standards and therapeutic strategies, while a separate table compares in-context learning with two-phase prompting.Source: Table 10 · p. 18, Table 11 · p. 18
• Figures 10–14 show OpenR1-Psy prompts for interaction planning, dialogue reconstruction, DSM/ICD-grounded empathy, validation, and diagnostic classification.Source: Figure 10 · p. 19, Figure 11 · p. 19, Figure 12 · p. 20
• Figures 15–17 show prompts for data evaluation, performance evaluation, and internal-thinking evaluation.Source: Figure 15 · p. 22, Figure 16 · p. 23, Figure 17 · p. 24
• Table 12 lists the criteria for counseling-quality assessment.Source: Table 12 · p. 25