Words Caught in the Heart: Simulating Person-Specific Expression for Self-Opaque Clients in Psychotherapy

Words Caught in the Heart. Simulating Person-Specific Expression for Self-Opaque Clients in Psychotherapy. By Mengxi Xiao and fourteen others, from Wuhan University, the Cognitive Language and Information Lab, and four other institutions. Published on the thirty-first of October.

People in psychotherapy can feel distress well before they are able to put it into coherent words. How they express themselves shifts with emotion, with how safe they feel, and with trust. Client simulators built on large language models often start from clinically organized case descriptions, and that starting point can push them to explain the client's difficulties more clearly than the client's current understanding would actually allow. The simulated client ends up sounding like the case file.

Pandora is a client simulation framework built around a different question: how psychological experience turns into expression that is situated and individual. Rather than reading a clinical summary aloud, the simulated client works from its own evolving state, and a dedicated planning step decides how that state should come out in words.

As the therapist intervenes, Pandora updates the client's psychological and relational states. It constructs subjective appraisals and attributions, and it plans both what the response should contain and what strategy it should take. A separate expression planner then combines those decisions with language patterns distilled from the client's own utterances, and uses them to guide how the reply is realized linguistically. Content and voice are planned apart, then brought together.

Pandora is compared with five baselines on one hundred and fifteen cases drawn from four counseling and motivational interviewing datasets. It reaches the highest expression similarity and the highest overall disclosure-alignment score among the systems compared. In a study with four evaluators over forty-five matched disclosure items, Pandora's responses were judged AI-like in seven point two percent of ratings.

The baselines drew that judgment in forty-three point three percent of ratings at best, and up to sixty-seven point two percent at worst. Removing the expression planner raised the figure to thirty-five point six percent, and the same ablation also lowered expression similarity.

The authors read these results as support for planning expression explicitly in client simulation, as a way to match individual language patterns and to sound less like a machine, within the setting they evaluated. The ablation carries much of that weight: the planner is the piece whose removal moves both numbers in the wrong direction.