Expression planning cuts AI-like judgments in psychotherapy client simulation from 43% to 7.2%

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

Mengxi Xiao, Jingjing Chen, Ziyan Kuang, Jiarong Song, Wenhao Zhao, Zuchen Han, Yihuai Xu, Xiyang Huang, Jie Gong, Yuechen Jiang, Xueqing Peng, Yankai Chen, Xiaoyan Qu, Min Peng, Qianqian Xie

cs.HC

2026-10-31

Pandora adds an explicit expression-planning stage to LLM psychotherapy client simulation: top expression similarity on 115 cases, 7.2% AI-like judgments vs 43.3–67.2% for five baselines.

What problem this solves

LLM-based client simulators already exist for training and evaluating therapists: Patient-Ψ pairs CBT case formulations with style labels, MindVoyager models metacognition and openness, PatientAct connects clinical and relational patterns. They share a weakness: generation is driven by clinically organized case descriptions that assemble experiences, beliefs, and coping into a coherent interpretation, so the simulated client explains difficulties more clearly than their current understanding warrants. The title's word for this is self-opaque clients, people who feel distress before they can describe it.

How much a real client can say varies with emotional intensity, perceived safety, and trust; a concern often surfaces first as a concrete incident or a tentative remark. For training this is a practical risk: an unusually articulate simulated client makes a difficult concern easier to identify than it would be in practice. A style label shifts tone while leaving the psychological explanation oddly explicit. The problem, in short: the content is right and the manner of speaking is not.

Method

Pandora borrows Levelt's distinction between conceptual preparation and linguistic formulation, splitting what to say from how to say it into two explicit pipelines, five steps per turn:

One design choice carries weight: case material is time-truncated and fully visible to the simulator, while what the client currently understands is a separate, evolving quantity. Visibility belongs to the system; understanding belongs to the client.

Results

115 cases (97 Chinese, 18 English) from PsyDial-D4, AnnoMI, Eeyore Full, and ConsistentMI, against five baselines and an ablation without expression planning:

MethodGDS↑Sstyle↑AIR↓
Pandora0.8710.5667.2%
Pandora w/o expression planning0.8280.45835.6%
Consistent MI (best baseline)0.6640.51247.2%
Interactive Agents0.6430.50143.3%
OpenClaw0.6090.39761.7%

AIR comes from a human study: 45 matched disclosure items (15 per memory level), four evaluators (two counselors, two synthetic-data researchers), binary AI-like judgments, 180 per source. Pandora drew 13 AI-like votes (7.2%) against 43.3–67.2% for the baselines, a 36.1-point drop from Interactive Agents. No Pandora candidate received a majority (3 of 4 or more) of AI votes; 35 of 45 received none. The reference dialogues scored 4.4%.

The ablation is the most informative number: removing expression planning leaves GDS at 0.828, still above every baseline, but drops content-reduced similarity from 0.566 to 0.458, below Consistent MI, and raises AIR to 35.6%, a paired difference of 28.3 points (case-bootstrap 95% CI 20.6–36.1). Disclosure organization survives; the individual voice does not. On the 97 Chinese cases Pandora's Sstyle is 0.489 against 0.427 for Consistent MI, so the advantage holds where the metric fits the language best. Layer-wise timing alignment leads at every level (0.843/0.854/0.866). Interactive Agents keeps the lowest bundling-frequency error (0.099 vs 0.102).

Why it matters

For anyone building user or persona simulation, the headline evidence is architectural: semantic planning and linguistic realization deserve separate modules. Upstream stages can organize content well and still fail to sound like a specific person; an explicit realization interface delivers both similarity and human-perception gains at once, and nothing in it is specific to psychotherapy. The code is released. For mental-health tooling, the paper names a measurement blind spot: a simulator given away by AI-like phrasing distorts training scenarios. The claim stays where the paper leaves it: what is validated is expression fidelity and perceived AI-likeness, not benefit to therapist training.

Limitations

From the authors: the case pool is predominantly Chinese and includes reconstructed or generated material; the numerical state settings are design choices; the ablation evaluates expression planning as a whole rather than individual signals; the human study uses four evaluators, all co-authors, rating isolated utterances with modest agreement (mean pairwise 62.9%) and unlogged source-reveal, so blinding is incomplete. Reading closely: the Core115 benchmark and most metrics (LDDA, GDS, the AIR protocol) were built for this paper, with no external benchmark for comparison; the content-masking lexicon is Chinese-oriented, so English-case evidence is coarse; the reference AIR of 4.4% sits near Pandora's 7.2%, but the reference itself contains generated material, so this is not proof of human-level realism. Interactive Agents still wins bundling-frequency error, suggesting short direct replies have their own strengths.

Terms

Source

Related papers

All paper explainers