PsychAgent-Qwen3-32B: An Advanced AI Psychological Counseling Model
ecnu-icalk's PsychAgent-Qwen3-32B is a 32 billion parameter model based on Qwen/Qwen3-32B, specifically engineered for multi-session psychological counseling. Unlike traditional static models, PsychAgent employs an experience-driven lifelong learning framework to enhance longitudinal consistency and counseling quality.
Key Capabilities & Innovations
- Memory-Augmented Planning Engine (MAPE): Maintains evolving client profiles and session summaries for continuous, session-level planning.
- Skill Evolution Engine (SEE): Extracts and organizes therapeutic skills into a hierarchical tree based on practice.
- Reinforced Internalization Engine (RIE): Internalizes successful counseling strategies through rejection fine-tuning, making useful approaches endogenous.
- Longitudinal Consistency: Designed to improve consistency across multiple counseling sessions.
- High Performance: Outperforms general-purpose and psychology-specific baselines on all four aggregated PsychEval dimensions, including Counselor Shared, Counselor Specific, Client Shared, and Client Specific metrics. It also ranks first in human evaluation across Ethics, Interaction, Intervention, and Perception.
- Context Length: Trained with a maximum context length of 32,768 tokens.
Ideal Use Cases
- Research: Perfect for research into AI psychological counseling, longitudinal dialogue agents, and lifelong learning for counseling.
- Experimentation: Suitable for experiments involving memory, planning, and skill evolution in counseling agents.
- Benchmarking: Excellent for benchmarking in multi-session counseling settings, particularly those similar to PsychEval.
Important Note: This model is not a substitute for licensed mental health professionals and should not be used for clinical care, emergencies, or crisis intervention. Its observed improvements are benchmark trends, not direct clinical evidence.