A newly posted survey on Hugging Face examines how large language models (LLMs) are being applied to mental health care, arguing that the field has moved through three increasingly sophisticated phases. The authors frame their review around the growing global burden of mental health conditions and persistent barriers in traditional care, including limited resources, high costs, stigma, and privacy concerns.
According to the survey, Phase I treats LLMs primarily as passive information tools and pattern recognizers used for assessment. Phase II shifts toward empathetic conversationalists capable of in-the-moment but stateless interactions. The current frontier, Phase III, aims at longitudinal, personalized companions built as stateful cognitive agents.
To support this framework, the authors systematically review core technologies, agent architectures — organized around profile, memory, reasoning, and planning — and the datasets and benchmarks that underpin the field's development. They note that the rapidly expanding body of work has lacked a coherent evolutionary narrative, making it difficult to contextualize progress and identify future directions.
The survey positions itself as a synthesis of existing work and a roadmap for responsible, effective, and human-centered AI in mental healthcare. A curated collection of the reviewed resources is available at the project's GitHub repository.