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

Pandora simulates psychotherapy clients by separating psychological state and content planning from an individual expression planner, improving expression similarity and reducing perceived AI-like judgments in the reported evaluation.

Existing LLM-based client simulators often rely on clinically organized case descriptions, which can encourage clients to explain difficulties more coherently than their current understanding warrants. Clients may seek help before they can coherently describe distress, and their expression can vary with emotional distress, perceived safety, and trust. Conversational style labels may alter tone while leaving the client's psychological explanation unusually explicit.

Pandora takes the therapist utterance, dialogue history, previous psychological state, case material, and an individual language profile as inputs at each turn. A rule-based intervention evaluator identifies clarification, questioning, and emotional-reflection opportunities and produces an intervention signal; bounded rule-based transitions update affective, relational, defensive, and reflective state dimensions. Subjective appraisal and attribution are constructed from the updated state, case material, and intervention signal to represent what the experience means and how the client currently explains it. A content plan with stance and a response-act plan specify what to communicate and the immediate conversational purpose, conditioned on the therapist utterance. A separate expression planner combines the state, content plan, history, and offline language profile to produce realization guidance, which the client LLM uses to generate the utterance.

• Pandora achieves the highest expression similarity and overall disclosure-alignment score among the compared systems on 115 cases from four counseling and motivational interviewing datasets. §Abstract
• In Table 2, Pandora scores 0.852 LDDA, 0.875 DS, 0.871 GDS, 0.566 S_style, and 0.385 S_raw, versus 0.811, 0.795, 0.828, 0.458, and 0.311 for Pandora w/o EP. §4.3 Experimental Results
• Pandora receives AI-like judgments in 7.2% of ratings, compared with 43.3–67.2% for the baselines and 35.6% without expression planning. §Abstract
• The final collection contains 115 cases: 97 Chinese and 18 English, comprising 97 PsyDial-D4, 11 AnnoMI, five Eeyore Full, and two ConsistentMI cases. §4.1 Data Collection and Case Selection
• The human evaluation includes 45 matched disclosure items and four evaluators, yielding 180 judgments per source and 1,440 judgments in total. §E.3 Endpoint Aggregation and Uncertainty

• Stated in Limitations: the case collection is predominantly Chinese and includes reconstructed or generated material; broader coverage of languages, cultures, and independently collected client speech remains to be evaluated.
• Stated in Limitations: psychological state variables and their numerical settings are design choices, and the current ablation evaluates expression planning as a whole rather than isolating finer-grained conditioning signals.
• Stated in Limitations: human evaluation uses a small panel and isolated utterances, with variation across raters; source-reveal usage was not logged, so complete blinding cannot be confirmed.
• Stated in the Ethical Statement: the measures do not establish clinical safety or treatment efficacy, and internal psychological states are modeling constructs rather than clinical findings about real individuals.

• The paper identifies a gap between clinically organized client representations and clients' own expression, framing simulation around how psychological experience becomes situated speech. §1 Introduction
• Pandora connects psychological state updates with subjective appraisal and attribution, content and response strategy planning, and a separate expression plan. §3 Method
• The state separates affective, relational, defensive, and reflective dimensions, allowing willingness to disclose and capacity to articulate to vary separately. §3.1 Therapist Intervention Evaluation and State Update
• Attribution ranges from immediate behavior and circumstances to tentative relational interpretations and connections with underlying beliefs, conditioned by state and intervention signal. §3.2 Subjective Appraisal and Attribution
• Response strategy maps factual questions, practical requests, reflections, and challenges or invalidation to different immediate conversational purposes. §3.3 Content Target, Client Stance, and Response Strategy
• Expression planning combines a stable individual language profile distilled offline from client utterances with current state, content plan, and dialogue history. §3.4 Individual Expression Planning
• The ablation retains the preceding psychological and content planning stages while removing the expression module. §4.3 Experimental Results
• Human evaluators were four co-authors: two qualified counselors with over 1,000 hours of counseling experience each and two synthetic-data researchers. §E.2 Evaluator Qualifications and Interface