Agentic Stage-Based LLM Framework for Multi-Turn Mental Health Support Conversations in Thai
2
Issued Date
2025-01-01
Resource Type
Scopus ID
2-s2.0-105032722447
Journal Title
2025 20th International Joint Symposium on Artificial Intelligence and Natural Language Processing Isai Nlp 2025
Rights Holder(s)
SCOPUS
Bibliographic Citation
2025 20th International Joint Symposium on Artificial Intelligence and Natural Language Processing Isai Nlp 2025 (2025)
Suggested Citation
Changpun R., Khoprasertthaworn N., Jongpipatchai P., Petcharat T., Rungsimontuchat K., Suttiwan P., Nupairoj N., Hemrungrojn S., Tuicomepee A., Achakulvisut T., Vateekul P. Agentic Stage-Based LLM Framework for Multi-Turn Mental Health Support Conversations in Thai. 2025 20th International Joint Symposium on Artificial Intelligence and Natural Language Processing Isai Nlp 2025 (2025). doi:10.1109/iSAI-NLP66160.2025.11320484 Retrieved from: https://repository.li.mahidol.ac.th/handle/123456789/115798
Title
Agentic Stage-Based LLM Framework for Multi-Turn Mental Health Support Conversations in Thai
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
While mental health support needs are growing in Thailand, access to professional support remains limited. LLMbased chatbots offer a scalable solution, yet most existing systems are restricted to single-turn, solution-oriented responses. We propose an Agentic Stage-Based LLM Framework for Multi- Turn Mental Health Support Conversations in Thai, structuring conversations around five core counseling stages: rapport building, problem identification, goal setting, working, and termination. Drawing from Person-Centered Therapy (PCT) and Acceptance and Commitment Therapy (ACT), our framework incorporates two key components: an approach-selection agent that selects appropriate counseling approaches and a monitoring agent that manages stage transitions in multi-turn conversations. We evaluated the proposed framework against three single-agent baselines using LLM-simulated users and compared positive user reaction rates as judged by LLMs. Our framework achieved a 79.01% positive user reaction rate, outperforming single agents with standard, AugESC, and COOPER-CoT prompts by 8.91%, 11.68%, and 17.02%, respectively. Ablation studies validated the necessity of each agentic module in the proposed framework, with results surpassing single-agent baselines without stage-based architecture by 96.15% and 88.46% in Process and Working evaluations, respectively. A/B testing with real users and evaluation by 3 counseling practitioners demonstrated significant improvements in seven of eight mental health support evaluation metrics, highlighting the potential to deliver LLM-based mental health support in Thai.