Paper review · November 5, 2026
Words That Catch: Simulating Individual Expression in Psychotherapy Clients with Limited Self-Understanding
Words Caught in the Heart: Simulating Person-Specific Expression for Self-Opaque Clients in Psychotherapy
The authors built Pandora, a client simulator with an "expression planning" stage separate from psychological state, subjective interpretation and content planning, improving the reproduction of individual speech styles and reducing AI-like expressions.
Problem
LLM-based client simulators used for counselor training and evaluation often rely on clinically organized case descriptions. As a result, they tend to explain difficulties too clearly, even ones a real client has not yet put into words. Real clients express themselves hesitantly or in fragments depending on their emotions, sense of safety and trust.
Method
Following theories of speaker language production, Pandora plans "what to say" and "how to say it" separately. First, a rule-based evaluator judges the counselor's intervention (whether clarification, questions or emotional reflection were used), and updates the emotional, relational, defensive and reflective states accordingly. Next, it constructs subjective appraisals and attributions, and sets a content plan (stance: hesitation, defensiveness, etc.) and a response strategy. Finally, it creates an expression plan that combines the current state with a linguistic profile extracted offline from the client's utterances (discourse markers, hedging, emotional explicitness, length distribution), and the LLM follows this plan to generate the utterance.
Results
- The authors compared Pandora with five baseline systems on 115 cases selected from four counseling and motivational interviewing datasets. Pandora had the highest expression similarity and global disclosure alignment score (GDS).
- Four raters (two counselors and two synthetic data researchers) judged 45 disclosure items, and Pandora was judged "AI-like" 7.2% of the time. The baseline systems ranged from 43.3% to 67.2%.
- In an ablation without expression planning, the AI-like rate rose to 35.6% and expression similarity also dropped. The authors present this as evidence that the expression planning stage contributes.
Limitations
Of the 115 cases, 97 are in Chinese (PsyDial-D4). The masking lexicon used for content-reduction similarity is also Chinese-centric, so the evidence for English cases is weaker. The human evaluation involves four raters and 45 items. The authors state that the automatic metrics do not demonstrate psychological authenticity or clinical equivalence. State transitions depend on rule-based numeric settings.
Why it matters
If a simulator for counselor training produces clients who are too articulate, practice becomes easier than reality. This work shows that user simulation needs to explicitly plan individual speech style, not just content. It is also worth consulting when designing user simulators in other domains.
Key terms
- Client simulator
- A system that makes an LLM play the role of a client receiving psychotherapy, for counselor training or for evaluating counseling AI.
- Expression planning
- A stage that decides in advance how to express a given response content in the person's speech style (hedging, emotional explicitness, length, etc.).
- Subjective appraisal and attribution
- A separate representation of what meaning the client takes from an experience (appraisal) and how they currently explain it (attribution). It is distinct from a clinician's case conceptualization.
- Ablation
- An experiment that removes one component of a system and measures the change in performance to determine that component's contribution.
Original
Authors: Mengxi Xiao, Jingjing Chen, Ziyan Kuang, Jiarong Song, Wenhao Zhao, Zuchen Han, Yihuai Xu, Xiyang Huang, et al. · Posted 2026-10-31 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.