From Pattern Recognizers to Personalized Companions: A Survey of Large Language Models in Mental Health

From Pattern Recognizers to Personalized Companions. A Survey of Large Language Models in Mental Health. By He Hu and ten others. Published on the twenty-first of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Mental health conditions are rising in prevalence worldwide, and traditional healthcare has long faced barriers: limited resources, high cost, stigma, and concerns about privacy. Together these create an urgent need for support that is accessible and can scale. Large Language Models have emerged as a technology with strong potential to widen access to mental health support, through advanced natural language understanding and generation. Language is the medium most mental health support already runs through, which is why a technology built on it is seen as such a fit. But the body of work is growing fast and in pieces, and it lacks a coherent story of how it has developed. That makes it hard to put current progress in context or to see where the field should go next. The survey sets out to supply that missing story.

The survey's central thesis is that the role of LLMs in mental health is evolving through three distinct phases, each more sophisticated than the last. In Phase I, LLMs act mainly as passive information tools and pattern recognizers, used for assessment. In Phase II, they become empathetic conversationalists, handling interactions in the moment that carry no state from one conversation to the next.

Phase III is the current frontier. It seeks longitudinal, personalized companions, built as stateful cognitive agents that persist and adapt over time. The move from Phase II to Phase III is a move from remembering nothing to remembering the person. Read as a sequence, the three phases describe a steady increase in what the model is asked to be. First a tool that reads signals. Then a partner in a single conversation. Then a presence that persists across many conversations and adapts to one individual.

To support the framework, the survey systematically reviews the core technologies, agent architectures, and the infrastructure of datasets and benchmarks. The architectures are broken into four parts: Profile, Memory, Reasoning, and Planning. For each area it shows how its evolution underpins the developmental path from one phase to the next. The four architectural parts line up naturally with the third phase. A companion that persists needs a profile of the person, a memory of past interactions, the ability to reason about them, and the ability to plan ahead.

Seen through this developmental lens, the authors offer a comprehensive synthesis of existing work, a narrative of its trajectory, and a roadmap for future work on AI for mental healthcare that is responsible, effective, and human-centred.

The organising idea is a progression from recognising patterns, to holding a conversation, to keeping someone company over time. It gives a fragmented literature a single axis to line up along. Any given system can be placed by asking which phase it belongs to, and what it would need to reach the next. That is also where the roadmap comes from: the gaps between the current frontier and the third phase are the work still to do. A curated collection of the resources reviewed is available in the project repository.