PaperlayerView exact revision on arXiv

Source-linked AI summary

From Pattern Recognizers to Personalized Companions: A Survey of Large Language Models in Mental Health

He Hu, Yucheng Zhou, Qianning Wang, Yingjian Zou, Chiyuan Ma, Juzheng Si, Jianzhuang Liu, Zitong Yu, Laizhong Cui, Fei Ma, Qi Tian

arXiv:2609.25186v1cs.CYcs.AIcs.CL

TL;DR

Barriers to traditional mental healthcare accompany rising mental health needs, while prior surveys have treated narrower application areas or broad generative-AI applications. This survey organizes the literature into three phases, from assessment tools to stateless conversationalists and stateful personalized companions, and synthesizes technologies, architectures, datasets, and benchmarks. It concludes that the field is striving toward Phase III agents with memory, planning, and tool use.

  • Problem

    Mental health needs are rising amid barriers to traditional care, and prior surveys focus on narrower areas or lack this survey's three-phase framing.

    Source: §I · p. 1

  • Method

    The survey organizes and reviews mental health LLM work through a three-phase framework and synthesizes core technologies, agent architectures, datasets, and benchmarks.

    Source: §I · p. 2

  • Results

    The field has moved from passive risk-detection tools through supportive single-session dialogue and is now pursuing stateful companions with memory, planning, and tool use.

    Source: §VII · pp. 13–14

  • Takeaways & Limitations

    The survey links the field's trajectory to evolving technologies, agent architectures, datasets, and benchmarks as a basis for future directions.

    Source: §I · p. 2

Abstract

from arXiv · show

The rising global prevalence of mental health conditions, together with longstanding barriers in traditional healthcare, such as limited resources, high cost, stigma, and privacy concerns, has created an urgent need for accessible and scalable support. Large Language Models (LLMs) have emerged as a transformative technology with strong potential to democratize mental health support through advanced natural language understanding and generation. However, the rapidly expanding, fragmented body of work in this area lacks a coherent evolutionary narrative, making it difficult to contextualize current progress and identify future directions. This survey addresses this gap by organizing and analyzing the literature around a central thesis: the role of LLMs in mental health is evolving through three distinct, increasingly sophisticated phases. We trace this trajectory from Phase I, in which LLMs act primarily as passive Information Tools and Pattern Recognizers for assessment; through Phase II, where they function as Empathetic Conversationalists for in-the-moment, stateless interactions; to the current frontier, Phase III, which seeks Longitudinal, Personalized Companions implemented as stateful cognitive agents. To support this framework, we systematically review core technologies, agent architectures (Profile, Memory, Reasoning, and Planning), and the critical infrastructure of datasets and benchmarks, highlighting how their evolution underpins this developmental path. Viewing the field through this developmental lens, we provide a comprehensive synthesis of existing work, an insightful narrative of its trajectory, and a clear roadmap for future innovation in responsible, effective, and human-centered AI for mental healthcare. A curated collection of the resources reviewed in this survey is available at our project repository: https://github.com/Emo-gml/Awesome-Mental-Health-LLMs.

I. INTRODUCTION

This survey frames mental health LLM research as an evolution from passive assessment tools to conversational support and stateful personalized companions. It connects that trajectory to technical and architectural developments, datasets and benchmarks, and future research directions.

I. INTRODUCTION

  • Traditional mental healthcare faces resource scarcity, high costs, stigma, and privacy concerns, while rising mental health needs create demand for accessible support.Source: §I · p. 1
  • The survey presents three phases: static-text assessment by Information Tools, stateless in-session support by Empathetic Conversationalists, and a frontier of Longitudinal, Personalized Companions.Phase I is described as reactive pattern recognition; Phase II supports coherent, empathetic dialogue within an isolated session.Source: §I · pp. 1–2
  • Phase III calls for stateful agents with memory for past interactions, dynamic user profiles, and goal-oriented planning across sessions.These components are proposed to support long-term therapeutic relationships.Source: §I · p. 2
  • The survey links the phases to reviews of core technologies, agent architectures, datasets, and benchmarks, and outlines challenges and future directions.It also positions its synthesis as more comprehensive than prior surveys focused on narrower application areas.Source: §I · pp. 1–3

II. BACKGROUND AND APPLICATION

The section surveys four mental health application areas and presents them as part of a shift from passive pattern recognition toward active conversational support. It then introduces criteria and adaptation methods for developing mental health LLMs.

II. BACKGROUND AND APPLICATION

  • Across four application pillars, LLMs progress from passive pattern recognition toward active conversational partnerships that lay groundwork for Phase III agents.The pillars are assessment and diagnosis, therapeutic and clinical interventions, education, and professional assistance.Source: §II · p. 3

A. Mental Health Assessment and Diagnosis

  • LLMs analyze social-media and forum text for mental health indicators, help score PHQ-9 and GAD-7, and use multimodal cues for assessment.The surveyed tasks include depression, anxiety, and suicide-risk identification; multimodal models incorporate speech prosody and facial expressions.Source: A. Mental Health Assessment and… · p. 3

B. Therapeutic and Clinical Interventions

  • LLMs support open-ended, context-aware therapeutic dialogue, including CBT-inspired interventions and everyday emotional companionship.Multimodal signals can provide tone and nonverbal behavior for interactions described as more like human counseling.Source: B. Therapeutic and Clinical… · p. 3

C. Education

  • LLMs simulate virtual patients with varied symptoms and communication styles so trainees can practice interviewing, therapeutic communication, and crisis management.These simulations offer repeated, controlled exposure without relying solely on scarce human resources.Source: C. Education · p. 3

D. Professional Assistance and Clinical Tools

  • LLMs can automate session summaries, extract clinical themes, and help formulate treatment plans, streamlining clinician workflows and freeing more attention for patient care.Source: D. Professional Assistance and… · p. 3

III. LARGE LANGUAGE MODELS FOR MENTAL HEALTH

  • The survey uses interaction granularity, user-state modeling, memory, personalization, and autonomy to distinguish systems across the three phases.It presents these criteria as a lens for understanding the technical developments discussed in the section.Source: §III · p. 3, Table I · p. 4

A. Core Methodologies for Model Adaptation

  • Adapting general-purpose LLMs for mental health requires changing their knowledge, behavior, and values through a multi-stage training and alignment process.The section frames these methods around clinical validity, empathy, and safety.Source: A. Core Methodologies for Model… · p. 4, §III · p. 4

1) Domain-Specific Pre-Training:

The survey traces mental-health LLM specialization from domain knowledge and therapeutic dialogue training to inference-time reasoning and grounding. These techniques support more clinically aligned interactions, while data quality and transfer remain important constraints.

1) Domain-Specific Pre-Training:

  • Domain-specific pretraining supplies mental-health knowledge that general web pretraining lacks, supporting more robust downstream risk assessment.The survey notes that pretrained speech models may entangle acoustic and semantic information, limiting downstream depression detection.Source: 1) Domain-Specific Pre-Training · p. 4

2) Supervised Fine-Tuning (SFT) for Therapeutic Competence:

  • Supervised fine-tuning on curated counseling dialogues teaches therapeutic patterns, emotional nuance, and conversation structure for Phase II.The survey distinguishes full fine-tuning from parameter-efficient fine-tuning.Source: 2) Supervised Fine-Tuning (SFT) for… · p. 4
  • Full fine-tuning adapts all parameters to psychological discourse, whereas PEFT trains lightweight modules while freezing most parameters.Full fine-tuning can deeply embed therapeutic logic [17],, [55]; PEFT methods such as LoRA and QLoRA target domain adaptation with lower computational cost,,.Source: 2) Supervised Fine-Tuning (SFT) for… · p. 4
  • PEFT supports diverse therapeutic expertise and practical deployment, but SFT can constrain flexibility, amplify dataset bias, and impair out-of-distribution transfer.Examples span CBT, Narrative Therapy, and Socratic reasoning; the survey also notes on-device fine-tuning under privacy constraints.Source: 2) Supervised Fine-Tuning (SFT) for… · p. 4

3) Alignment with Reinforcement Learning:

  • Reinforcement learning refines SFT-trained behavior toward safe, helpful, therapeutic interactions using simulated feedback or reasoning-quality rewards.Preference methods include DPO, KTO, and ORPO,,; Psyche-R1 uses GRPO to improve reasoning on difficult cases [88], [89].Source: 3) Alignment with Reinforcement Learning · p. 4

1) Prompt Engineering for Cognitive Alignment:

  • Prompt engineering uses instructions, examples, or psychological constructs to steer clinical reasoning without parameter updates.Prompts can incorporate validated instruments such as the BDI to guide interpretable, evidence-based explanations.Source: 1) Prompt Engineering for Cognitive… · pp. 4–5

2) Chain-of-Thought for Reasoning Transparency; 3) Retrieval-Augmented Generation (RAG) for Safety and Grounding:

  • Chain-of-Thought prompting exposes intermediate reasoning, while RAG retrieves clinical knowledge or user history to ground decisions and adapt support to context.CoT systems draw on therapeutic theories [17],, and are reported to improve diagnostic reliability and expert alignment,,,; context-sensitive RAG can condition tone and focus on emotional state or intent,,,,.Source: 2) Chain-of-Thought for Reasoning… · p. 5, 3) Retrieval-Augmented Generation… · p. 5

C. Beyond Text: Integrating Multimodal Cues

Multimodal mental-health systems combine text with acoustic and visual cues for assessment and empathetic intervention, but interpreting those cues requires caution. The survey then frames the move from stateless conversations toward agents with user models and persistent capabilities.

C. Beyond Text: Integrating Multimodal Cues

1) Multimodal Systems for Empathetic Intervention:

  • Empathetic-intervention systems combine dialogue with facial, bodily, speech, or video cues to recognize emotion and shape responses, yet cues can reflect dataset artifacts rather than intent.Empatheia incorporates text, speech, and avatar video; higher multimodal performance does not necessarily imply more clinically appropriate empathy.Source: 1) Multimodal Systems for Empathetic… · p. 5

2) Multimodal Systems for Clinical Assessment:

  • Clinical-assessment systems fuse audio-visual signals and text-derived features for depression severity or PHQ-8 prediction, but context-dependent cues can introduce bias and spurious correlations.The survey calls for validation across diverse populations and settings.Source: 2) Multimodal Systems for Clinical… · p. 5

D. Transitioning from Episodic Conversations to Longitudinal Companionship

  • Phase II systems can produce clinically relevant, empathetic single-session dialogue, but they remain stateless and lack persistent user understanding.The survey presents stateful cognitive architectures as the shift toward Phase III longitudinal companionship.Source: D. Transitioning from Episodic… · p. 5

IV. LLM-BASED AGENT FOR MENTAL HEALTH

  • Mental-health LLM agents are designed to emulate cognitive abilities such as memory, reasoning, and planning to support safe, reliable tools.The section examines longitudinal personalization, clinical frameworks, multi-agent reasoning, and responsible development.Source: §IV · pp. 5–6

A. System Architecture and Core Components

1) Profile:

  • The Profile module maintains a changing user model, including emotional states, psychological indicators, and task-specific traits, to tailor support.The survey connects structured profiles with tailored interventions and consistent reactions to individual stressors [21].Source: 1) Profile · p. 6

2) Memory:

The survey describes agent components for continuity, clinical reasoning, planning, and tool access, then connects these capabilities to long-term engagement and clinically guided interaction. Together, these approaches shift attention from recalling dialogue toward managing a therapeutic journey.

2) Memory:

  • Memory enables cross-session recall through hierarchical stores, dialogue retrieval, or psychologically inspired mechanisms such as forgetting-curve simulation and cognitive-distortion tracking.Examples include hierarchical short- and long-term structures,, RAG over historical dialogue, and specialized mechanisms,.Source: 2) Memory · p. 6

3) Reasoning:

  • Reasoning modules analyze user input, form clinical judgments, and develop response strategies through structured workflows, multi-perspective deliberation, and explainable traces.These methods aim to align agent logic with professional diagnostic and intervention standards [104].Source: 3) Reasoning · p. 6

4) Planning:

  • Planning makes interactions goal-oriented by decomposing long-term objectives into sessions, aligning dialogue with therapeutic stages, or steering topics.The survey describes this as giving multi-turn interactions an overarching therapeutic purpose.Source: 4) Planning · p. 6

5) Tool Use:

  • Tool use grounds responses in external clinical resources, with some systems conditioning retrieval on speech timing or emotional signals.Retrieved resources include guidelines, CBT scripts, and crisis helpline directories,,,,,.Source: 5) Tool Use · p. 6

B. Enhancing Longitudinal Engagement and Personalization

  • Longitudinal engagement requires modeling relational dynamics and therapeutic goals, not only retrieving prior dialogue.MusPsy models gradual client-state evolution, CA+ [21] aligns sessions with long-term goals, and iPET uses a virtual-pet paradigm to foster attachment.Source: B. Enhancing Longitudinal Engagement… · p. 6

C. Integrating Clinical Frameworks and Therapeutic Modalities

  • Clinical frameworks enter agents through imitation of specialized dialogue, explicit procedural guidance, and reasoning over clinical skills and theories.The survey describes synthetic CBT dialogues, staged narrative therapy, planned Motivational Interviewing, and reasoning about cognitive distortions, or DSM and ICD criteria [17].Source: C. Integrating Clinical Frameworks and… · pp. 6–7

D. Leveraging Multi-Agent Architectures for Complex Reasoning and Interaction

The survey presents multi-agent systems as a way to divide complex counseling work, alongside safety, evaluation, and data infrastructure for responsible development. It also traces datasets from static labels toward longitudinal and synthetic interactions.

D. Leveraging Multi-Agent Architectures for Complex Reasoning and Interaction

  • Multi-agent architectures divide counseling tasks because a single LLM must otherwise combine empathy, diagnosis, and intervention planning.The survey groups decomposition across cognitive, procedural, and social dimensions as a pathway toward more powerful mental-health agents.Source: D. Leveraging Multi-Agent… · p. 7
  • Collaborating agents can debate interpretations, handle stages of structured clinical workflows, or simulate multiple human perspectives.Examples include agents assigned CBT phases, specialized agents implementing MINI interview branching, and virtual social-support roles [112].Source: D. Leveraging Multi-Agent… · p. 7

E. Ensuring Safety, Reliability, and Responsible Development

  • Safety mechanisms act at different levels—shaping policy, filtering outputs, or grounding facts—and automated safety gains may still miss context-sensitive clinical risks.The survey identifies false reassurance, delayed escalation, and overconfident pseudo-clinical reasoning, and calls for evaluation incorporating clinical context, longitudinal dynamics, and risk-sensitive criteria [163], [164], [165].Source: E. Ensuring Safety, Reliability, and… · p. 7
  • Evaluation should measure clinical utility and safety beyond single-turn helpfulness, calibrating LLM-as-a-Judge against experts and validated clinical constructs.ψ-Arena uses simulated-client, agent, and supervisor feedback; IMA tracks narrative change, while PHQ-9 provides a clinical basis for assessments.Source: E. Ensuring Safety, Reliability, and… · pp. 7–8
  • Synthetic and simulated resources address privacy constraints and support training, but risk mismatch with real populations, model artifacts, and cultural homogenization.The survey describes theory-guided synthetic datasets, multi-session datasets from anonymized case reports,, and simulated patients for therapist training,, [171].Source: E. Ensuring Safety, Reliability, and… · p. 8

V. DATASETS

  • Dataset development mirrors the field’s progression from static classification labels to single-session dialogues and longitudinal synthetic agent interactions.The survey identifies data synthesis as a current bridge toward Phase III.Source: §V · p. 8

A. Resources for Phase I: Detection and Risk Assessment

Phase I resources center on static classification, while Phase II datasets shift toward therapeutic dialogue and empathy; real-world counseling data offers ecological validity but remains constrained by privacy.

A. Resources for Phase I: Detection and Risk Assessment

  • Phase I datasets were predominantly static social-media classification resources, including depression and suicide-risk detection and datasets for comorbidity, severity, and clinical risk stratification.Recent resources broadened coverage to stress, cognitive distortions, and self-harm intent, with some adding explanatory annotations.Source: A. Resources for Phase I: Detection… · pp. 8–9
  • Foundational Phase I resources established ground-truth data for detection and risk assessment.The passage identifies RSDD, UMD, and DR as foundational resources.Source: A. Resources for Phase I: Detection… · p. 8

B. Resources for Phase II: Therapeutic Dialogue and Empathy

2) Real-World Counseling Transcripts:

  • Real-world counseling datasets offer high ecological validity and support analysis of therapist strategies and client reactions, but privacy limits their scale and often reduces them to single-turn or anonymized material.These constraints hinder modeling long-term therapeutic trajectories.Source: 2) Real-World Counseling Transcripts · p. 10

C. Resources for Phase III: Longitudinal and Agent-Centric Data

Phase III faces a data famine because real-world datasets rarely capture the continuous, multi-session changes needed to train memory and planning. Emerging resources target longitudinal continuity and structured clinical reasoning.

C. Resources for Phase III: Longitudinal and Agent-Centric Data

  • Real-world data rarely captures the continuous, multi-session evolution of a patient’s state needed to train memory and planning modules.This gap is described as a “data famine” for the Phase III frontier.Source: C. Resources for Phase III:… · p. 10
  • MusPsy, DiaCBT, and PsyDial are designed to track patients across sessions or long-term conversations, supporting learning of long-range contextual dynamics.These datasets target longitudinal continuity.Source: C. Resources for Phase III:… · p. 10
  • ESD-CoT, OpenR1-Psy, Psyche-R1, StimuliQA, and Psy-Insight provide reasoning or explainability annotations that expose clinical logic behind responses.Psy-Insight is described as providing rich explainable annotations.Source: C. Resources for Phase III:… · p. 10

D. Paradigm Shift: Advanced Data Synthesis

Data synthesis progresses from multi-turn dialogue to psychologically structured and personalized-realism scenarios, while multimodal resources broaden available signals. Benchmarking is also shifting toward clinical and dynamic evaluation, though synthetic data and model-based judgments retain important limitations.

D. Paradigm Shift: Advanced Data Synthesis

  • Synthetic counseling data progresses from multi-turn dialogue expansion to psychologically structured content and simulations conditioned on user profiles and therapeutic goals.The three levels are dialogue-level, cognitive-level, and personalized-realism synthesis.Source: D. Paradigm Shift: Advanced Data… · pp. 10–11
  • Synthetic dialogues can lack longitudinal consistency, realistic emotional dynamics, and structural fidelity, creating risks including distributional shift, echo-chamber effects, and cultural homogenization.The text contrasts locally coherent generated dialogue with the nonlinear patterns of real therapy.Source: D. Paradigm Shift: Advanced Data… · p. 11

E. Multimodal Resources

  • Multimodal resources combine visual, acoustic, and physiological signals, supporting dialogue-interaction tasks and behavior-based state detection.Examples include audio-visual datasets for empathy and CBT-skill assessment and resources for stress and depression detection.Source: E. Multimodal Resources · p. 11

VI. BENCHMARKS

  • Mental-health evaluation is moving from text matching toward clinical simulations, but LLM-as-a-Judge remains a proxy rather than clinically grounded assessment and may have limited reliability and expert agreement.The proposed evaluation emphasis includes longitudinal consistency, escalation behavior, crisis handling, and therapeutic alliance.Source: §VI · pp. 11–12

A. Safety and Compliance Benchmarks

  • Safety benchmarks test harmful-content avoidance, crisis-resource provision, and robustness to adversarial or ambiguous inputs.SafeBench, PsyCrisis-Bench, and CHBench assess high-risk scenarios; PRiMH uses adversarial testing.Source: A. Safety and Compliance Benchmarks · p. 12

B. Clinical Competency and Process Evaluation

  • Clinical-competency benchmarks assess adherence to therapeutic techniques, diagnostic reliability, and relational constructs such as client-belief inference and therapeutic alliance.The surveyed resources include CBT-Bench, CounselingBench, ESC-Judge, diagnosis benchmarks, and broader suites such as MentalChat16K.Source: B. Clinical Competency and Process… · p. 12

C. Frontier: Dynamic Simulation and Agent Evaluation

Phase III evaluation is moving beyond static question-answering toward simulated longitudinal interactions. The proposed metric hierarchy progresses from surface overlap to dialogue quality and clinical or relational attributes, with model-based judgments requiring caution.

C. Frontier: Dynamic Simulation and Agent Evaluation

  • Dynamic evaluation uses simulated clients to test Phase III agents’ longitudinal capabilities, including long-term consistency and adaptive planning.Interactive platforms use profiled Patient Agents in multi-turn or multi-session dialogues, and some frameworks combine simulated-client feedback, LLM supervisors, and objective outcomes.Source: C. Frontier: Dynamic Simulation and… · p. 13

1) Tier 1: Automated Surface Metrics:

  • BLEU, ROUGE, F1, and Distinct-N support reproducibility but correlate poorly with therapeutic quality because they measure lexical overlap and fluency rather than empathy or clinical insight.Source: 1) Tier 1: Automated Surface Metrics · p. 13

2) Tier 2: Semantic and Dialogue Quality:

3) Tier 3: Clinical and Relational Attributes:

  • Clinical and relational evaluation includes empathy, therapeutic alliance, efficacy, helpfulness, safety, and professionalism, with LLM judges serving as an imperfect proxy for clinical judgment.LLM-judge reliability varies, calibration is unclear, and agreement with expert judgments is limited; clinically grounded evaluation is more reliable but costly and difficult to scale.Source: 3) Tier 3: Clinical and Relational… · p. 13

E. Evaluation Gaps in Mental Health LLM Systems

Current evaluation relies mainly on offline proxies, with real-patient clinical validation remaining uncommon and constrained. The survey proposes multi-session, clinically grounded evaluation for Phase III and situates it within the progression toward longitudinal companions.

E. Evaluation Gaps in Mental Health LLM Systems

  • Real-patient clinical validation remains exceptional, leaving most systems without robust evidence of effectiveness under real-world therapeutic conditions.Table VII summarizes representative systems' validation settings, distinguishing expert, clinical, and automatic evaluation.Source: E. Evaluation Gaps in Mental Health… · p. 13, Table VII · p. 14

F. Evaluation Protocols for Phase III Systems

  • Phase III evaluation should assess four dimensions: longitudinal consistency, escalation behavior, crisis handling, and therapeutic alliance.These criteria extend evaluation beyond single-turn response quality to multi-session and clinically grounded assessment.Source: F. Evaluation Protocols for Phase III… · p. 13
  • Validated clinical instruments can connect automated metrics with clinical assessment and support trajectory-level benchmarking of mental health agents.The survey names CTRS, PANAS, and WAI as measures for therapeutic quality and fidelity, emotional dynamics, and relational quality.Source: F. Evaluation Protocols for Phase III… · p. 13

VII. CONCLUSION AND FUTURE DIRECTIONS

  • The field has progressed from passive risk-detection tools through single-session conversationalists toward stateful companions with memory, planning, and tool use.The survey says the transition to fully autonomous, clinically valid agents remains incomplete.Source: §VII · pp. 13–14

A. From Conversationalists to Cognitive Agents

The survey identifies persistent memory and planning, clinical alignment, simulation-based evaluation, and human-AI collaboration as directions for developing longitudinal mental-health agents. It frames these agents as augmenting clinicians, with high-risk cases escalated to professionals.

A. From Conversationalists to Cognitive Agents

  • Longitudinal agents need memory that stores and updates user information over time, plus proactive planning for interventions spanning multiple sessions.The survey contrasts this cognitive-architecture direction with context-window-based recall and reactive response generation.Source: A. From Conversationalists to… · p. 14

B. Clinical Alignment and Reasoning Transparency

  • Clinical alignment should prioritize established treatment protocols and transparent clinical judgments that permit human oversight.The survey suggests CoT for explainability and RAG linked to validated clinical manuals for evidence-based therapies.Source: B. Clinical Alignment and Reasoning… · p. 14

C. Revolutionizing Evaluation: The Simulation Paradigm

  • High-fidelity patient simulations can stress-test therapist agents before deployment and measure clinical outcomes beyond linguistic metrics.Suggested outcomes include simulated symptom reduction, such as PHQ-9 scores, and therapeutic-alliance improvement measured with WAI.Source: C. Revolutionizing Evaluation: The… · p. 14

D. Human-AI Symbiosis and Ethical Deployment

  • AI should augment clinicians through routine monitoring, psychoeducation, and between-session support, while escalating complex or high-risk cases to professionals.The survey also calls for robustness to cultural bias and nuanced handling of ethical dilemmas.Source: D. Human-AI Symbiosis and Ethical… · p. 14

Related papers

Loading 2609.25186v1…