Research map / Multi-agent systems and games / LLM agents

AgentMental: An Interactive Multi-Agent Framework for Explainable and Adaptive Mental Health Assessment

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 3 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AgentMental, a multi-agent framework that simulates clinical doctor-patient interviews with adaptive follow-up questioning and a tree-structured memory to score PHQ-8 depression items on DAIC-WOZ.

Problem

Automated mental health assessment mostly relies on static text analysis of transcripts, which cannot handle incomplete or vague user responses or integrate multi-topic counseling information into a coherent assessment.

Approach

Multi-agent system (question generator, evaluation, scoring, updating agents) that conducts interactive interviews, uses an evaluation agent's necessity score to trigger up to d follow-up questions per topic, and stores extracted information in a tree-structured memory (user, topic, statement nodes) that is updated across topics.

Key results

• On DAIC-WOZ with Qwen2.5-72B, AgentMental reaches MAE 2.514, Kappa 79.8, and Macro F1 89.8, versus best baseline MDAgents (sl) Kappa 73.6 and Macro F1 86.7.
• With Qwen2.5-14B, AgentMental achieves Kappa 72.4 and Macro F1 86.0, above the best 14B baseline (Chain-of-Logic / MDAgents, Macro F1 about 83.8 under sl).
• Ablation shows removing both in-depth questioning and memory drops Kappa from 79.8 to 47.2 and raises MAE from 2.514 to 3.400.

Surprising findings

• Item-level F1 on psychomotor changes remains low (28.9) even for the proposed method, which the authors attribute to subtle contextual cues.

Limitations

• Evaluation relies on LLM-simulated patients (DeepSeek-R1-Distill-Qwen-32B) rather than real participants in the interactive setting.
• Human and GPT-4o evaluation covers only 5 participant cases.
• Item-level performance is weak on subjective items such as psychomotor changes.

Experimental setup

Datasets: DAIC-WOZ

Benchmarks: DAIC-WOZ

Metrics: MAE Cohen's Kappa F1[C] F1[D] Macro F1 Understanding Sympathy Consistency Realism Coherence

Baselines: Zero-Shot 5-Shot CoT Chain-of-Logic Majority Voting Debate MDAgents Multi-View SEGA SEGA++

Builds on: AutoGen MDAgents chain-of-thought PHQ-8

Closest papers by meaning

• MoodAngels: A Retrieval-augmented Multi-agent Framework for Psychiatry Diagnosis (NeurIPS 2025 · similarity 0.88)
• MedEvoEval: Evaluating Continual Evolution of Doctor Agents through Simulated Clinical Episodes (NeurIPS 2026 · similarity 0.85)
• An LLM-based Simulation Framework for Embodied Conversational Agents in Psychological Counseling (AAAI 2026 · similarity 0.85)
• MedAgentBoard: Benchmarking Multi-Agent Collaboration with Conventional Methods for Diverse Medical Tasks (NeurIPS 2025 · similarity 0.85)
• Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models (NeurIPS 2025 · similarity 0.84)
• MDD-5k: A New Diagnostic Conversation Dataset for Mental Disorders Synthesized via Neuro-Symbolic LLM Agents (AAAI 2025 · similarity 0.84)