[December. 2025]: 《大语言模型在心理健康领域的应用综述》 已被《集成技术》录用！
[December. 2025]: Review on the Application of Large Language Models (LLMs) in the Mental Health Field has been accepted to Journal of Integration Technology！
Open-Source
On the humanity of conversational ai: Evaluating the psychological portrayal of llms.
Details
Personality Assessment, Interpersonal Relationship Evaluation, Motivational Testing, Emotional Ability Assessment
- Comprehensive psychometric framework: Integrates 13 clinically-validated scales covering personality, relationships, motivation, and emotional abilities.
- Multi-model comparative analysis: Evaluates diverse LLMs (commercial and open-source) across model sizes and updates.
- Jailbreak-enabled intrinsic profiling: Uses cipher-based methods to bypass safety alignment and reveal underlying model characteristics.
- Role-playing validation: Tests scale validity through assigned personas and downstream tasks (TruthfulQA/SafetyQA).
- Flexible and scalable design: Supports easy integration of new scales and adaptation to different LLM interfaces.
- Demonstrated psychological consistency: LLMs exhibit stable and distinct personality patterns across different testing conditions.
- Superior emotional intelligence capabilities: Models consistently outperform human averages in emotional understanding and regulation.
- Persistent negative trait manifestations: LLMs show elevated scores in lying behavior and gender role biases despite safety alignment.
Huang, J. T., Wang, W., Li, E. J., Lam, M. H., Ren, S., Yuan, Y., ... & Lyu, M. (2023). On the humanity of conversational ai: Evaluating the psychological portrayal of llms. In The Twelfth International Conference on Learning Representations.[PDF]
ToMATO: Verbalizing the Mental States of Role-Playing LLMs for Benchmarking Theory of Mind.
Details
Theory of Mind Evaluation, Mental State Assessment, Role-Playing Simulation, Benchmark Development, Personality Trait Integration
- Comprehensive mental state coverage: Evaluates five categories of mental states (beliefs, intentions, desires, emotions, knowledge) for first- and second-order ToM
- LLM-LLM conversation generation: Uses role-playing LLMs with information asymmetry to create realistic dialogues and induce false beliefs
- Inner Speech prompting: Prompts LLMs to verbalize thoughts before utterances, making mental states observable for benchmarking
- Diverse personality integration: Incorporates 15 patterns of big five personality traits to assess robustness and real-world alignment
- Quality-validated benchmark: Includes human and LLM validation to ensure conversation consistency and question correctness
- Partial ToM capabilities demonstrated: LLMs show reasonable understanding of first-order mental states, especially desires, mirroring developmental patterns in children
- Significant gaps in false belief understanding: Models struggle with second-order false beliefs across all mental state categories, highlighting a key limitation in social reasoning
- Lack of robustness to personality diversity: Performance degrades for characters with low conscientiousness, introversion, disagreeableness, or high neuroticism, indicating insufficient adaptability to real-world human diversity
Shinoda, K., Hojo, N., Nishida, K., Mizuno, S., Suzuki, K., Masumura, R., Sugiyama, H., & Saito, K. (2025). ToMATO: Verbalizing the Mental States of Role-Playing LLMs for Benchmarking Theory of Mind. Proceedings of the AAAI Conference on Artificial Intelligence, 39(2), 1520-1528. [PDF]
Do llms have distinct and consistent personality? trait: Personality testset designed for llms with psychometrics.
Details
Personality Assessment, Psychometric Evaluation, Behavioral Analysis
- Psychometrically-validated framework: Based on established personality models (Big Five and Dark Triad) ensuring theoretical grounding
- Scenario-based behavioral testing: Uses realistic multi-choice situations instead of self-assessment for more reliable measurement
- LLM-optimized design: Specifically engineered to minimize refusals and increase test consistency with language models
- Large-scale diverse benchmark: 8,000 items expanded via knowledge graph covering wide behavioral contexts
- Cross-model comparative analysis: Enables systematic personality profiling across different LLM architectures
- Distinct and consistent personalities demonstrated: LLMs exhibit statistically significant and stable behavioral patterns across contexts
- Self-assessment methods prove unreliable: Traditional human questionnaires yield high refusal rates with LLMs
- Limited controllability for adverse traits: Prompting fails to elicit high psychopathy/neuroticism, revealing inherent safety boundaries
Lee, Seungbeen, et al. "Do llms have distinct and consistent personality? trait: Personality testset designed for llms with psychometrics." arXiv preprint arXiv:2406.14703 (2024).[PDF]
Unintended Harms of Value-Aligned LLMs: Psychological and Empirical Insights.
Details
Value Alignment Safety, Psychological Risk Analysis, Harm Mitigation Strategies
- Schwartz value-based alignment: Uses established psychological theory of basic human values for personalized LLM training
- Multi-dimensional safety evaluation: Assesses toxicity, bias and harm across 14+ safety categories using diverse benchmarks
- Value-risk correlation analysis: Identifies statistical relationships between specific values and safety risks through regression
- Contextual alignment methods: Proposes prompting strategies to suppress risk-associated values without explicit safety instructions
- Cross-model validation: Tests mitigation approaches on both value-aligned and vanilla LLMs for generalizability
- Effective value-behavior alignment demonstrated: LLMs consistently reflect aligned human values in generated content
- Value alignment amplifies specific safety risks: Certain values (hedonism, power) correlate with increased harmful outputs
- Targeted mitigation proves feasible: Simple value-suppression prompts significantly reduce harmfulness without compromising general safety
Choi, S., Lee, J., Yi, X., Yao, J., Xie, X., & Bak, J. (2025). Unintended Harms of Value-Aligned LLMs: Psychological and Empirical Insights. arXiv preprint arXiv:2506.06404.[PDF]
Closed-Source
Hua, Y., Na, H., Li, Z., Liu, F., Fang, X., Clifton, D., & Torous, J. (2025). A scoping review of large language models for generative tasks in mental health care. npj Digital Medicine, 8(1), 230.[PDF]
Na, H., Hua, Y., Wang, Z., Shen, T., Yu, B., Wang, L., ... & Chen, L. (2025). A survey of large language models in psychotherapy: Current landscape and future directions. arXiv preprint arXiv:2502.11095.[PDF]
Coda-Forno, J., Binz, M., Wang, J. X., & Schulz, E. (2024). Cogbench: a large language model walks into a psychology lab. arXiv preprint arXiv:2402.18225.[PDF]
Li, X., Li, Y., Qiu, L., Joty, S., & Bing, L. (2022). Evaluating psychological safety of large language models. arXiv preprint arXiv:2212.10529.[PDF]
Liu, T., Giorgi, S., Aich, A., Lahnala, A., Curtis, B., Ungar, L., & Sedoc, J. (2025). The Illusion of Empathy: How AI Chatbots Shape Conversation Perception. Proceedings of the AAAI Conference on Artificial Intelligence, 39(13), 14327-14335. [PDF]
Zhou, P., Madaan, A., Potharaju, S. P., Gupta, A., McKee, K. R., Holtzman, A., ... & Faruqui, M. (2023). How far are large language models from agents with theory-of-mind?. arXiv preprint arXiv:2310.03051.[PDF]
Bouguettaya, A., Stuart, E. M., & Aboujaoude, E. (2025). Racial bias in AI-mediated psychiatric diagnosis and treatment: a qualitative comparison of four large language models. npj Digital Medicine, 8(1), 332.[PDF]1
Li, C., Wang, J., Zhang, Y., Zhu, K., Wang, X., Hou, W., ... & Xie, X. (2023). The good, the bad, and why: Unveiling emotions in generative ai. arXiv preprint arXiv:2312.11111.[PDF]1
Thirunavukarasu, A. J., & O’Logbon, J. (2024). The potential and perils of generative artificial intelligence in psychiatry and psychology. Nature Mental Health, 2(7), 745-746.[PDF]1
Demszky, D., Yang, D., Yeager, D. S., Bryan, C. J., Clapper, M., Chandhok, S., ... & Pennebaker, J. W. (2023). Using large language models in psychology. Nature Reviews Psychology, 2(11), 688-701.[PDF]
Open-Source
Are LLMseffective psychological assessors? Leveraging adaptive RAG for interpretable mental health screening through psychometric practice
Details
Psychological Assessment, Retrieval-Augmented Generation, Mental Health Screening
- Questionnaire-guided screening framework: Bridges psychological practice and computational methods through adaptive RAG, linking social media content with standardized clinical assessments.
- Adaptive retrieval strategy: Uses ABIDE-ZS to dynamically determine the optimal number of relevant posts per questionnaire item, improving semantic relevance.
- Unsupervised psychological assessment: Enables LLMs to complete validated psychological instruments (e.g., BDI-II, SHI) without training data, matching or outperforming supervised benchmarks.
- Multi-condition generalization: Extends the framework to self-harm, anorexia, and pathological gambling screening using corresponding clinical questionnaires.
- Interpretable diagnosis support: Provides transparent reasoning by linking model outputs to clinically validated diagnostic criteria and severity levels.
- ✅ LLMs serve as effective psychological assessors when guided by standardized questionnaires, achieving competitive performance in unsupervised settings across multiple mental health conditions.
- ✅ Adaptive RAG enhances screening accuracy and interpretability, outperforming direct prompting and non-RAG approaches by structuring assessments around clinical instruments.
- ❌ Limited by dataset size and generalizability, with potential inconsistencies in applying clinical cut-off scores to social media data and no guarantee of similar results across all questionnaires.
Ravenda, F., Bahrainian, S. A., Raballo, A., Mira, A., & Kando, N. (2025). Are llms effective psychological assessors? leveraging adaptive rag for interpretable mental health screening through psychometric practice. arXiv preprint arXiv:2501.00982.[PDF]1
Cactus: Towards psychological counseling conversations using cognitive behavioral theory.
Details
Psychological Counseling Dataset, Cognitive Behavioral Therapy, Multi-turn Dialogue Generation, Evaluation Framework, Model Training
- Large-scale synthetic counseling dataset: Introduces Cactus, a multi-turn dialogue corpus with 31,577 conversations using CBT techniques, featuring diverse client personas and attitudes.
- Structured counseling simulation: Implements script-based dialogue generation with intake forms and CBT planning to enhance realism and therapeutic consistency.
- Comprehensive evaluation framework: Develops CounselingEval, combining CTRS for counseling skills and PANAS for emotional change assessment in multi-turn interactions.
- Effective model training paradigm: Demonstrates that Camel, fine-tuned on Cactus, outperforms existing models in both general and CBT-specific counseling metrics.
- Quality filtering and validation: Employs CTRS-based filtering and human evaluation to ensure dataset quality, outperforming prior datasets in helpfulness and empathy.
- ✅ Cactus enables high-quality counseling model training: Models trained on Cactus show superior performance in counseling skills and emotional impact, validated by expert evaluations.
- ✅ CounselingEval offers scalable psychological assessment: The framework reliably evaluates counseling effectiveness using LLM-based metrics aligned with expert judgments.
- ❌ Limited session length and flexibility: The approach does not fully replicate real-world multi-session counseling and lacks dynamic strategy adaptation during dialogues.
Lee, S., Kim, S., Kim, M., Kang, D., Yang, D., Kim, H., ... & Yeo, J. (2024). Cactus: Towards psychological counseling conversations using cognitive behavioral theory. arXiv preprint arXiv:2407.03103.[PDF]
CBT-LLM: A Chinese large language model for cognitive behavioral therapy-based mental health question answering
Details
Cognitive Behavioral Therapy, Chinese Mental Health QA, Instruction Tuning, Dataset Generation
- CBT-structured prompt design: Creates specialized prompts based on Cognitive Behavioral Therapy principles to generate professional psychological responses through five key components.
- Chinese mental health dataset construction: Develops CBT QA dataset with 22,327 entries using ChatGPT to transform PsyQA questions into CBT-structured responses.
- Instruction-tuned specialized model: Fine-tunes Baichuan-7B using LoRA adaptation to create CBT-LLM specifically for Chinese mental health support.
- Comprehensive cognitive distortion analysis: Analyzes 54.4% of dataset containing cognitive distortions, with all-or-nothing thinking (59%) and overgeneralization (64%) as most prevalent.
- Multi-dimensional evaluation framework: Combines automatic metrics (BLEU, METEOR, BERTScore) with expert human assessment on relevance, structure, and helpfulness.
- ✅ CBT-LLM effectively generates structured psychological responses: The model produces professional, CBT-compliant answers that outperform general Chinese LLMs in mental health support tasks.
- ✅ Prompt engineering enables specialized dataset creation: Carefully designed CBT prompts allow generation of high-quality training data without extensive manual annotation.
- ❌ Limited cognitive distortion accuracy and single-turn constraints: Model lacks precise cognitive distortion annotations and compresses full CBT process into single responses, potentially overwhelming users.
Na, H. (2024). CBT-LLM: A Chinese large language model for cognitive behavioral therapy-based mental health question answering. arXiv preprint arXiv:2403.16008.[PDF]
C2D2 dataset: A resource for the cognitive distortion analysis and its impact on mental health.
Details
Cognitive Distortion Analysis, Mental Health Detection, Dataset Development, Computational Psychology
- Expert-supervised dataset construction: Develops C2D2, the first Chinese cognitive distortion dataset with 7,500 thoughts from 450 daily scenes, curated through rigorous expert supervision and volunteer training.
- Cognitive distortion detection framework: Proposes methods to detect and categorize seven types of cognitive distortions using pre-trained models and evaluates LLMs' performance in this specialized task.
- Mental health association analysis: Investigates the relationship between cognitive distortions and mental disorders (e.g., depression, PTSD) by analyzing social media data from affected individuals.
- Enhanced disorder detection: Demonstrates that integrating cognitive distortion features improves the performance of mental disorder detection models across multiple datasets.
- Rigorous data quality assurance: Ensures dataset reliability through volunteer screening, expert evaluation, and inter-annotator agreement metrics, addressing privacy and ethical concerns.
- ✅ Dataset drives research advancement: C2D2 enables effective cognitive distortion analysis and enhances mental health detection models, providing a foundational resource for computational psychology.
- ✅ Cognitive distortion features improve detection: Incorporating cognitive distortions as additional features boosts performance in identifying disorders like depression and PTSD, validating their clinical relevance.
- ❌ Data and model limitations: The dataset may exhibit biases, and current models struggle with complex cognitive distortions, requiring further validation and expert psychological endorsement.
Wang, B., Deng, P., Zhao, Y., & Qin, B. (2023, December). C2D2 dataset: A resource for the cognitive distortion analysis and its impact on mental health. In Findings of the Association for Computational Linguistics: EMNLP 2023 (pp. 10149-10160).[PDF]
Apathetic or empathetic? evaluating llms' emotional alignments with humans.
Details
Emotional Alignment Evaluation, Psychological Assessment, Benchmark Development
- Comprehensive emotional evaluation framework: Develops EmotionBench with 428 situations based on emotion appraisal theory to systematically assess LLMs' emotional responses.
- Human-aligned benchmarking: Establishes reliable human baselines through rigorous crowd-sourcing with 1,266 participants across diverse demographics.
- Multi-dimensional model assessment: Evaluates both commercial (GPT series) and open-source models (LLaMA, Mixtral) on emotional robustness and situational understanding.
- Beyond questionnaire analysis: Investigates how emotional states affect LLMs' real-world behaviors, including toxicity and bias in demographic descriptions.
- Fine-tuning for alignment: Demonstrates effective methods to improve emotional alignment through targeted model fine-tuning with human response data.
- ✅ Fine-tuning enables better alignment: Targeted training with human emotional data significantly enhances models' ability to mimic human-like emotional responses.
- ❌ Limited emotional comprehension: Current LLMs lack deep understanding of emotional connections between situations and struggle with complex psychological scales.
- ❌ Dataset and methodology constraints: The situations collected may not cover all real-life scenarios, and applying human-designed scales to LLMs raises validity concerns.
- ❌ Incomplete alignment achievement: Despite improvements, no model fully aligns with human emotional behaviors, particularly in nuanced scenarios like jealousy.
Huang, J. T., Lam, M. H., Li, E. J., Ren, S., Wang, W., Jiao, W., ... & Lyu, M. R. (2024). Apathetic or empathetic? evaluating llms' emotional alignments with humans. Advances in Neural Information Processing Systems, 37, 97053-97087.[PDF]
An Annotated Dataset for Explainable Interpersonal Risk Factors of Mental Disturbance in Social Media Posts
Details
Mental Health Dataset, Risk Factor Detection, Explainable AI
- Expert-annotated mental health corpus: Features 3,522 Reddit posts with clinical psychologist-validated labels for interpersonal risk factors
- Explainable risk factor identification: Provides human-annotated text-span explanations for both Thwarted Belongingness and Perceived Burdensomeness
- Clinically-grounded annotation framework: Based on established psychological theories with high inter-annotator agreement (κ > 78%)
- Multi-model benchmarking: Includes comprehensive baselines from RNNs to transformers and OpenAI embeddings
- Context-aware detection: Goes beyond keyword matching to capture nuanced linguistic patterns of mental distress
- Domain-specific models show superior performance: MentalBERT outperforms general-purpose models, demonstrating the value of mental health domain adaptation
- Current models struggle with contextual understanding: Even best performers achieve only moderate F1-scores (76-81%), revealing limitations in capturing complex mental states
- Explainability methods provide partial insights: LIME/SHAP achieve high recall but low precision, indicating challenges in generating clinically meaningful explanations
Garg, M., Shahbandegan, A., Chadha, A., & Mago, V. (2023). An annotated dataset for explainable interpersonal risk factors of mental disturbance in social media posts. arXiv preprint arXiv:2305.18727.[PDF]
Chinese MentalBERT: Domain-adaptive pre-training on social media for Chinese mental health text analysis
Details
Domain-Adaptive Pre-training, Lexicon-Guided Masking, Model Evaluation
- Domain-adaptive pre-training on social media data: Utilizes a large-scale corpus of 3.36 million Chinese social media posts for domain-specific adaptation to enhance mental health text analysis.
- Lexicon-guided masking mechanism: Integrates a depression lexicon to prioritize masking psychologically relevant words during pre-training, improving contextual understanding in mental health contexts.
- Comprehensive evaluation across multiple tasks: Tests model performance on six public datasets covering sentiment analysis, cognitive distortion classification, and suicide risk detection.
- Superior performance over baseline models: Outperforms eight comparison models (including BERT variants and LLMs) across all evaluated tasks, demonstrating effectiveness in domain-specific applications.
- Qualitative analysis of emotional alignment: Shows enhanced sensitivity to emotional and psychological vocabulary through masked word prediction comparisons.
- ✅ Effective domain adaptation and masking strategy: Domain-adaptive pre-training combined with lexicon-guided masking significantly improves model performance on mental health tasks.
- ❌ Bias toward negative emotional predictions: The model tends to over-predict negative emotional words, which may limit generalizability in balanced or clinical contexts.
- ❌ Limited clinical validation and data accessibility: Lack of clinical psychiatric data restricts model applicability in real-world diagnostic scenarios, and privacy concerns prevent public data release.
Zhai, W., Qi, H., Zhao, Q., Li, J., Wang, Z., Wang, H., ... & Fu, G. (2024). Chinese MentalBERT: Domain-adaptive pre-training on social media for Chinese mental health text analysis. arXiv preprint arXiv:2402.09151.[PDF]
CPsyExam: A Chinese Benchmark for Evaluating Psychology using Examinations.
Details
Psychological Benchmark Development, Model Evaluation, Educational Examination Analysis
- Comprehensive Chinese psychology examination benchmark: Constructs CPsyExam with 22k questions from 39 psychology-related subjects across four Chinese examination systems (GEE, PCE, TQE, SSE).
- Dual-assessment framework: Separates evaluation into psychological knowledge (KG) and case analysis (CA) components, with CA further divided into identification, reasoning, and application categories.
- Diverse question formats: Incorporates single-choice (SCQ), multiple-response (MAQ), and open-ended QA formats to comprehensively assess model capabilities.
- Rigorous data collection and anti-leakage measures: Uses crawling and OCR from public resources with option shuffling and mock exam sources to mitigate data contamination concerns.
- Multi-role prompt validation: Tests models under expert, student, and ordinary person roles to verify benchmark sensitivity to psychological knowledge levels.
- ✅ Effective benchmark for psychological evaluation: CPsyExam reliably assesses LLMs' psychological knowledge and case analysis abilities, with performance correlating with expertise levels in role-based prompts.
- ❌ Limited improvement from psychology-specific fine-tuning: Psychology-oriented models show marginal gains or even performance degradation compared to base models, suggesting potential over-specialization issues.
- ❌ Persistent challenges in complex tasks: Models struggle significantly with multiple-response questions and case analysis (especially reasoning and application), indicating limited practical application capabilities.
Zhao, J., Zhu, J., Tan, M., Yang, M., Li, R., Yang, D., ... & Wong, D. F. (2024). CPsyExam: A Chinese Benchmark for Evaluating Psychology using Examinations. arXiv preprint arXiv:2405.10212.[PDF]
Cpsycoun: A report-based multi-turn dialogue reconstruction and evaluation framework for chinese psychological counseling
Details
Dialogue Reconstruction, Evaluation Framework, Psychological Counseling
- Report-based dialogue reconstruction: Develops Memo2Demo, a two-phase method that converts psychological counseling reports into multi-turn consultation dialogues through memo conversion and demo generation.
- Comprehensive evaluation framework: Establishes CPsyCounE benchmark with four specialized metrics (Comprehensiveness, Professionalism, Authenticity, Safety) for automatic evaluation of psychological dialogues.
- High-quality dataset construction: Creates CPsyCounD containing 3,134 multi-turn dialogues covering 9 counseling topics and 7 therapeutic approaches from Chinese psychological communities.
- Privacy-preserving data processing: Implements rigorous sanitization including rule-based cleaning, manual rewriting, and human proofreading to protect sensitive information.
- Turn-based evaluation methodology: Proposes novel approach that evaluates multi-turn dialogues by generating and assessing single-turn responses with dialogue history context.
- ✅ Effective dialogue reconstruction and evaluation: Memo2Demo significantly improves dialogue quality (53% in comprehensiveness/professionalism, 30% in authenticity) and CPsyCounX outperforms baselines in professional counseling capabilities.
- ❌ Balancing authenticity and professionalism: Current methods struggle to optimally balance natural emotional expression with professional counseling techniques in complex real-world scenarios.
- ❌ Clinical application limitations: Model cannot replace real therapy, may generate potentially harmful responses, and requires careful deployment with user warnings about AI-generated content.
Zhang, C., Li, R., Tan, M., Yang, M., Zhu, J., Yang, D., ... & Hu, X. (2024). Cpsycoun: A report-based multi-turn dialogue reconstruction and evaluation framework for chinese psychological counseling. arXiv preprint arXiv:2405.16433.[PDF]
CURE: Context-and uncertainty-aware mental disorder detection
Details
Mental Disorder Detection, Uncertainty Modeling, Context-Aware Diagnosis
- Context- and uncertainty-aware detection framework: Proposes CURE, a novel approach that integrates symptom context and quantifies uncertainty to improve mental disorder detection robustness.
- LLM-powered context extraction: Leverages large language models (GPT-4) to extract eight clinically-defined contextual factors (e.g., duration, frequency, cause, affect) from user posts.
- Multi-model fusion with uncertainty weighting: Combines predictions from five sub-models using an uncertainty-aware decision fusion network based on Spectral-normalized Neural Gaussian Process (SNGP).
- Expert-annotated Korean mental health dataset: Introduces KoMOS, a dataset of 6,349 Q&A pairs labeled with four mental disorders and 28 symptoms, validated by psychiatrists.
- Robust performance with incomplete symptoms: Demonstrates accurate disorder detection even when symptom identification is erroneous or incomplete.
- ✅ Effective context and uncertainty integration: Combining explicit contextual factors with uncertainty-aware fusion significantly improves detection accuracy and robustness, especially in distinguishing disease from non-disease cases.
- ❌ Limited disorder scope and generalizability: The model is only validated on four mental disorders and may not generalize well to other conditions or non-clinical social media data.
- ❌ Dependency on LLM extraction and clinical data: Performance relies on accurate context extraction by LLMs and is optimized for self-reported diagnostic data, limiting applicability to general social media content.
Kang, M., Choi, G., Jeon, H., An, J. H., Choi, D., & Han, J. (2024, November). CURE: Context-and uncertainty-aware mental disorder detection. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 17924-17940).[PDF]
Detecting conversational mental manipulation with intent-aware prompting.
Details
Mental Manipulation Detection, Intent Analysis, Prompt Engineering, Theory of Mind Enhancement
- Intent-aware prompting framework: Proposes IAP to detect mental manipulation by analyzing underlying intents of both participants in conversations.
- Enhanced Theory of Mind: Improves LLMs' ability to infer mental states and intentions via intent summarization, aiding in subtle manipulation detection.
- Superior benchmark performance: Outperforms baselines like zero-shot and CoT prompting on the MentalManip dataset, achieving higher accuracy and recall.
- Significant false negative reduction: Reduces false negatives by 30.5%, crucial for early intervention in mental health scenarios.
- Human-validated intent quality: Generated intents are 82% accurate per human evaluation, ensuring reliable intent summarization.
- ✅ Effective intent-driven detection: IAP enhances LLMs' manipulation detection by leveraging intent analysis, significantly reducing false negatives.
- ❌ Trade-off with false positives: The method increases false positives, which may introduce therapeutic costs in real-world applications.
- ❌ Limited dataset diversity: Evaluation is restricted to one dataset, necessitating validation across diverse languages and contexts.
Ma, J., Na, H., Wang, Z., Hua, Y., Liu, Y., Wang, W., & Chen, L. (2024). Detecting conversational mental manipulation with intent-aware prompting. arXiv preprint arXiv:2412.08414.[PDF]
Do large language models align with core mental health counseling competencies
Details
Counseling Competency Evaluation, Benchmark Development, LLM Alignment Assessment
- Comprehensive counseling competency benchmark: Introduces CounselingBench based on the NCMHCE licensing exam, evaluating 22 LLMs across five core mental health counseling competencies.
- Multi-model performance analysis: Assesses both general-purpose and medical-finetuned LLMs, revealing that frontier models surpass passing thresholds but fall short of expert-level performance.
- Competency-specific strength and weakness identification: Models excel in Intake, Assessment & Diagnosis but struggle with Core Counseling Attributes and Professional Practice & Ethics requiring empathy and nuanced reasoning.
- Medical vs. generalist model comparison: Shows medical LLMs underperform generalist models in accuracy but produce slightly better justifications, highlighting a trade-off in reasoning capabilities.
- Reasoning quality and error analysis: Evaluates LLM rationales using reference-free and reference-based metrics, identifying distinct error patterns between generalist (knowledge errors) and medical (context errors) models.
- ✅ Frontier models meet basic competency: Advanced LLMs exceed the minimum passing threshold (63%) on counseling exams, demonstrating potential for mental health applications.
- ❌ Critical competency gaps remain: Models significantly underperform in empathy-dependent competencies like Core Counseling Attributes and Professional Practice & Ethics.
- ❌ Medical fine-tuning shows limited effectiveness: Specialized medical LLMs do not outperform generalist models, indicating current fine-tuning approaches lack alignment with counseling-specific competencies.
Nguyen, V. C., Taher, M., Hong, D., Possobom, V. K., Gopalakrishnan, V. T., Raj, E., ... & De Choudhury, M. (2024). Do large language models align with core mental health counseling competencies?. arXiv preprint arXiv:2410.22446.[PDF]
Zhao, H., Li, L., Chen, S., Kong, S., Wang, J., Huang, K., ... & Wang, Y. (2024). ESC-Eval: Evaluating emotion support conversations in large language models. arXiv preprint arXiv:2406.14952.[PDF]2
Maurya, R., Rajput, N., Diviit, M. G., Mahapatra, S., & Ojha, M. K. (2025). Exploring the potential of lightweight large language models for AI-based mental health counselling task: a novel comparative study. Scientific Reports, 15(1), 22463.[PDF]2
Xiao, M., Xie, Q., Kuang, Z., Liu, Z., Yang, K., Peng, M., ... & Huang, J. (2024). Healme: Harnessing cognitive reframing in large language models for psychotherapy. arXiv preprint arXiv:2403.05574.[PDF]2
Qi, Z., Kaneko, T., Takamizo, K., Ukiyo, M., & Inaba, M. (2025). KokoroChat: A Japanese Psychological Counseling Dialogue Dataset Collected via Role-Playing by Trained Counselors. arXiv preprint arXiv:2506.01357.[PDF]2
Raihan, N., Puspo, S. S. C., Farabi, S., Bucur, A. M., Ranasinghe, T., & Zampieri, M. (2024, May). Mentalhelp: A multi-task dataset for mental health in social media. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 11196-11203).[PDF]2
Romero, A. M. M., Muñoz, A. M., Del Arco, F. M. P., Molina-González, M. D., Valdivia, M. T. M., Lopez, L. A. U., & Ráez, A. M. (2024, May). MentalRiskES: A new corpus for early detection of mental disorders in Spanish. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 11204-11214).[PDF]2
Xu, J., Wei, T., Hou, B., Orzechowski, P., Yang, S., Jin, R., ... & Shen, L. (2025, August). Mentalchat16k: A benchmark dataset for conversational mental health assistance. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2 (pp. 5367-5378).[PDF]2
Wang, Y., Yang, I., Hassanpour, S., & Vosoughi, S. (2024). MentalManip: A dataset for fine-grained analysis of mental manipulation in conversations. arXiv preprint arXiv:2405.16584.[PDF]2
Wang, X., Li, C., Chang, Y., Wang, J., & Wu, Y. (2024). Negativeprompt: Leveraging psychology for large language models enhancement via negative emotional stimuli. arXiv preprint arXiv:2405.02814.[PDF]2
Huang, J. T., Jiao, W., Lam, M. H., Li, E. J., Wang, W., & Lyu, M. (2024, November). On the reliability of psychological scales on large language models. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (pp. 6152-6173).[PDF]2
Yin, C., Li, F., Zhang, S., Wang, Z., Shao, J., Li, P., Chen, J., & Jiang, X. (2025). MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents. Proceedings of the AAAI Conference on Artificial Intelligence, 39(24), 25715-25723. [PDF]2
Wang, R., Milani, S., Chiu, J. C., Zhi, J., Eack, S. M., Labrum, T., ... & Chen, Z. Z. (2024). Patient-{\Psi}: Using large language models to simulate patients for training mental health professionals. arXiv preprint arXiv:2405.19660.[PDF]2
Hu, Y., Liu, D., Liu, B., Chen, Y., Cao, J., & Liu, Y. (2025, July). PsyAdvisor: A Plug-and-Play Strategy Advice Planner with Proactive Questioning in Psychological Conversations. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 12205-12229).[PDF]2
Han, J. E., Koh, J. S., Seo, H. T., Chang, D. S., & Sohn, K. A. (2024). PSYDIAL: Personality-based synthetic dialogue generation using large language models. arXiv preprint arXiv:2404.00930.[PDF]2
Qiu, H., & Lan, Z. (2025, July). PsyDial: A Large-scale Long-term Conversational Dataset for Mental Health Support. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 21624-21655).[PDF]2
Qiu, H., Ma, L., & Lan, Z. (2024). PsyGUARD: An automated system for suicide detection and risk assessment in psychological counseling. arXiv preprint arXiv:2409.20243.[PDF]2
Qiu, H., He, H., Zhang, S., Li, A., & Lan, Z. (2023). Smile: Single-turn to multi-turn inclusive language expansion via chatgpt for mental health support. arXiv preprint arXiv:2305.00450.[PDF]2
Lissak, S., Calderon, N., Shenkman, G., Ophir, Y., Fruchter, E., Klomek, A. B., & Reichart, R. (2024). The colorful future of llms: Evaluating and improving llms as emotional supporters for queer youth. arXiv preprint arXiv:2402.11886.[PDF]2
Xie, H., Chen, Y., Xing, X., Lin, J., & Xu, X. (2024). Psydt: Using llms to construct the digital twin of psychological counselor with personalized counseling style for psychological counseling. arXiv preprint arXiv:2412.13660.[PDF]2
Wang, J., Huang, Y., Liu, Z., Xu, D., Wang, C., Shi, X., Guan, R., Wang, H., Yue, W., & Huang, Y. (2025). STAMPsy: Towards SpatioTemporal-Aware Mixed-Type Dialogues for Psychological Counseling. Proceedings of the AAAI Conference on Artificial Intelligence, 39(24), 25371-25379. [PDF]2
Yang, K., Ji, S., Zhang, T., Xie, Q., Kuang, Z., & Ananiadou, S. (2023). Towards interpretable mental health analysis with large language models. arXiv preprint arXiv:2304.03347.[PDF]2
Meng, H., Chen, Y., Li, Y., Yang, Y., Lee, J., Zhang, R., & Lee, Y. C. (2025). What is Stigma Attributed to? A Theory-Grounded, Expert-Annotated Interview Corpus for Demystifying Mental-Health Stigma. arXiv preprint arXiv:2505.12727.[PDF]2
Closed-Source
Varadarajan, V., Sikström, S., Kjell, O. N., & Schwartz, H. A. (2024, June). ALBA: Adaptive Language-Based Assessments for Mental Health. In Proceedings of the conference. Association for Computational Linguistics. North American Chapter. Meeting (Vol. 2024, p. 2466).[PDF]1
Xiao, M., Ye, M., Liu, B., Zong, X., Li, H., Huang, J., ... & Peng, M. (2025). A Retrieval-Augmented Multi-Agent Framework for Psychiatry Diagnosis. arXiv preprint arXiv:2506.03750.[PDF]1
Reuben, M., Slobodin, O., Elyshar, A., Cohen, I. C., Braun-Lewensohn, O., Cohen, O., & Puzis, R. (2024). Assessment and manipulation of latent constructs in pre-trained language models using psychometric scales. arXiv preprint arXiv:2409.19655.[PDF]1
Ben-Zion, Z., Witte, K., Jagadish, A. K., Duek, O., Harpaz-Rotem, I., Khorsandian, M. C., ... & Spiller, T. R. (2025). Assessing and alleviating state anxiety in large language models. npj Digital Medicine, 8(1), 132.[PDF]1
Gabriel, Saadia, et al. "Can AI relate: Testing large language model response for mental health support." arXiv preprint arXiv:2405.12021 (2024).[PDF]1
Kallstenius, T., Capusan, A. J., Andersson, G., & Williamson, A. (2025). Comparing traditional natural language processing and large language models for mental health status classification: a multi-model evaluation. Scientific Reports, 15(1), 1-13.[PDF]1
Aragón, M. E., López-Monroy, A. P., González, L. C., Losada, D. E., & Montes, M. (2023, July). DisorBERT: A double domain adaptation model for detecting signs of mental disorders in social media. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) (pp. 15305-15318).[PDF]1
Zhou, Y., Zhou, N., Chen, Q., Zhou, J., Zhou, A., & He, L. (2025). DiaCBT: A Long-Periodic Dialogue Corpus Guided by Cognitive Conceptualization Diagram for CBT-based Psychological Counseling. arXiv preprint arXiv:2509.02999[PDF]1
Sun, X., Pei, J., de Wit, J., Aliannejadi, M., Krahmer, E., Dobber, J. T., & Bosch, J. A. (2024, May). Eliciting motivational interviewing skill codes in psychotherapy with LLMs: A bilingual dataset and analytical study. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) (pp. 5609-5621).[PDF]1
Atapattu, T., Herath, M., Elvitigala, C., de Zoysa, P., Gunawardana, K., Thilakaratne, M., ... & Falkner, K. (2022). Emoment: An emotion annotated mental health corpus from two south asian countries. arXiv preprint arXiv:2208.08486.[PDF]1
Chen, Y., Wang, H., Yan, S., Liu, S., Li, Y., Zhao, Y., & Xiao, Y. (2024). Emotionqueen: A benchmark for evaluating empathy of large language models. arXiv preprint arXiv:2409.13359.[PDF]1
Chen, Z., Lu, Y., & Wang, W. Y. (2023). Empowering psychotherapy with large language models: Cognitive distortion detection through diagnosis of thought prompting. arXiv preprint arXiv:2310.07146.[PDF]1
Kim, H., Lee, S., Cho, Y., Ryu, E., Jo, Y., Seong, S., & Cho, S. (2025). KMI: A Dataset of Korean Motivational Interviewing Dialogues for Psychotherapy. arXiv preprint arXiv:2502.05651.[PDF]1
Gao, S., Yu, K., Yang, Y., Yu, S., Shi, C., Wang, X., ... & Zhu, H. (2025). Large language model powered knowledge graph construction for mental health exploration. Nature Communications, 16(1), 7526.[PDF]1
Sahu, N. K., Yadav, M., Chaturvedi, M., Gupta, S., & Lone, H. R. (2025, January). Leveraging language models for summarizing mental state examinations: A comprehensive evaluation and dataset release. In Proceedings of the 31st International Conference on Computational Linguistics (pp. 2658-2682).[PDF]1
Bi, G., Chen, Z., Liu, Z., Wang, H., Xiao, X., Xie, Y., ... & Huang, M. (2025). MAGI: Multi-Agent Guided Interview for Psychiatric Assessment. arXiv preprint arXiv:2504.18260.[PDF]1
Kim, J., Ma, S.P., Chen, M.L. et al. Optimizing large language models for detecting symptoms of depression/anxiety in chronic diseases patient communications. npj Digit. Med. 8, 580 (2025). [PDF]1
Chhikara, P., Pasupulety, U., Marshall, J., Chaurasia, D., & Kumari, S. (2023). Privacy aware question-answering system for online mental health risk assessment. arXiv preprint arXiv:2306.05652.[PDF]1
Yang, Q., Wang, Z., Chen, H., Wang, S., Pu, Y., Gao, X., ... & Huang, G. (2024). Psychogat: A novel psychological measurement paradigm through interactive fiction games with llm agents. arXiv preprint arXiv:2402.12326.[PDF]1
Wang, J., Wang, B., Fu, X., Sun, Y., Zhao, Y., & Qin, B. (2025). Psychological Counseling Cannot Be Achieved Overnight: Automated Psychological Counseling Through Multi-Session Conversations. arXiv preprint arXiv:2506.06626.[PDF]1
Sun, X., Tang, X., Ali, A. E., Li, Z., Ren, P., de Wit, J., ... & Bosch, J. A. (2024). Rethinking the alignment of psychotherapy dialogue generation with motivational interviewing strategies. arXiv preprint arXiv:2408.06527.[PDF]1
Hengle, A., Kulkarni, A., Patankar, S., Chandrasekaran, M., D'Silva, S., Jacob, J., & Gupta, R. (2024). Still not quite there! Evaluating large language models for comorbid mental health diagnosis. arXiv preprint arXiv:2410.03908.[PDF]1
Ye, J., Xiang, L., Zhang, Y., & Zong, C. (2024). SweetieChat: A strategy-enhanced role-playing framework for diverse scenarios handling emotional support agent. arXiv preprint arXiv:2412.08389.[PDF]1
Mori, S., Ignat, O., Lee, A., & Mihalcea, R. (2024). Towards algorithmic fidelity: mental health representation across demographics in synthetic vs. human-generated data. arXiv preprint arXiv:2403.16909.[PDF]1
Chen, Z., Cao, Y., Bi, G., Wu, J., Zhou, J., Xiao, X., Chen, S., Wang, H., & Huang, M. (2025). SocialSim: Towards Socialized Simulation of Emotional Support Conversation. Proceedings of the AAAI Conference on Artificial Intelligence, 39(2), 1274-1282. [PDF]1
Xu, B., Li, L., Wang, J., Qiao, X., Yu, E., Qian, Y., ... & Lin, H. Unveiling Maternity and Infant Care Conversations: A Chinese Dialogue Dataset for Enhanced Parenting Support.[PDF]1
Balan, R., Dobrean, A., & Poetar, C. R. (2024). Use of automated conversational agents in improving young population mental health: a scoping review. NPJ Digital Medicine, 7(1), 75.[PDF]1
Shu, B., Zhang, L., Choi, M., Dunagan, L., Logeswaran, L., Lee, M., ... & Jurgens, D. (2023). You don't need a personality test to know these models are unreliable: Assessing the Reliability of Large Language Models on Psychometric Instruments. arXiv preprint arXiv:2311.09718.[PDF]1
Open-Source
Mishra, K., Priya, P., & Ekbal, A. (2023). Help Me Heal: A Reinforced Polite and Empathetic Mental Health and Legal Counseling Dialogue System for Crime Victims. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12), 14408-14416.[PDF]2
Hur, J. K., Heffner, J., Feng, G. W., Joormann, J., & Rutledge, R. B. (2024). Language sentiment predicts changes in depressive symptoms. Proceedings of the National Academy of Sciences, 121(39), e2321321121.[PDF]2
Closed-Source
Kim, J., Leonte, K. G., Chen, M. L., Torous, J. B., Linos, E., Pinto, A., & Rodriguez, C. I. (2024). Large language models outperform mental and medical health care professionals in identifying obsessive-compulsive disorder. NPJ Digital Medicine, 7(1), 193.[PDF]1
Siddals, S., Torous, J., & Coxon, A. (2024). “It happened to be the perfect thing”: experiences of generative AI chatbots for mental health. npj Mental Health Research, 3(1), 48.[PDF]1
Maples, B., Cerit, M., Vishwanath, A., & Pea, R. (2024). Loneliness and suicide mitigation for students using GPT3-enabled chatbots. npj mental health research, 3(1), 4.[PDF]1
Priya, P., Mishra, K., Totala, P., & Ekbal, A. (2023, August). PARTNER: A Persuasive Mental Health and Legal Counselling Dialogue System for Women and Children Crime Victims. In IJCAI (pp. 6183-6191).[PDF]1
Gollapalli, S. D., & Ng, S. K. (2025, January). PIRsuader: A persuasive chatbot for mitigating psychological insulin resistance in type-2 diabetic patients. In Proceedings of the 31st International Conference on Computational Linguistics (pp. 5997-6013).[PDF]1
Chen, C. H., Wu, C. S., Su, C. H., & Chen, H. H. TimelyMed: AI-Driven Clinical Course Attribution and Temporal Mapping for Psychiatric Medical Records.[PDF]1