Abstract: Existing Artificial Intelligence (AI) psychological counseling systems were generally limited to the counselor role, making it difficult to dynamically address users’ dual needs for both professional advice and emotional support. To overcome this limitation, a large-scale mental health language model named STAR (Supportive Therapeutic Adaptive Responder) was proposed, which was based on a dual-role interaction mechanism. First, a dynamic role-switching mechanism guided by system prompts was designed to enable transitions between the counselor and friend roles, addressing users’ dual needs in psychological counseling scenarios. Then, a training process incorporating data quality control was implemented, where model feedback was continuously collected to dynamically optimize both data generation strategies and system prompt design, thereby forming a closed loop between data generation and model training and enhancing dataset quality. Experimental results on the CPsyCoun (Chinese Psychological Counseling) benchmark show that the STAR model achieves improvements of 13.53% in comprehensiveness, professionalism, and authenticity over the base model Qwen2.5-7B-Instruct, and 26.81% over the single-role EmoLLM (Emotional Large Language Model). It also outperforms models fine-tuned on open-source datasets such as SoulChat, as well as general-purpose large models like GPT-4o-mini. These findings demonstrate that the STAR model effectively achieves dual-role adaptive responding, significantly enhancing both the effectiveness and experience of psychological counseling.

Key words: Large Language Model (LLM), psychological counseling, dual-role interaction, emotional support, instruction fine-tuning

摘要： 针对现有人工智能（AI）心理咨询系统仅扮演单一咨询师角色，难以动态适应用户对专业建议与情感支持的双重需求，本文提出了一种基于双角色交互机制的心理健康大语言模型STAR（Supportive Therapeutic Adaptive Responder）。首先，设计了基于系统提示的角色动态切换机制，实现心理咨询师与朋友角色的风格转换，满足用户在心理咨询场景中对专业建议和情感支持的双重需求。其次，采用数据质量控制的训练流程，持续收集模型反馈、动态优化数据生成策略与系统提示词设计，形成数据生成与模型训练的闭环，提升数据集质量。实验结果表明，在中文心理学基准测试框架CPsyCoun（Chinese Psychological Counseling）中，STAR模型在全面性、专业性与真实性上的综合得分相较于基座模型Qwen2.5-7B-Instruct提升了13.53%，较单角色心理大模型EmoLLM（Emotional Large Language Model）提升了26.81%，并优于使用灵心（SoulChat）等开源数据集微调的模型以及GPT-4o-mini等通用大模型。研究表明，STAR模型能够有效实现双角色自适应响应，有效提升心理咨询的效果和体验。

关键词: 大语言模型, 心理咨询, 双角色交互, 情感支持, 指令微调