Person-Centered Therapy (PCT)

• Person-Centered Therapy (PCT) is a humanistic approach emphasizing empathic understanding, unconditional positive regard, and congruence to foster client-led exploration.
• Recent computational research employs datasets like HamRaz and E-THER to operationalize PCT through structured prompt engineering and multimodal evaluation.
• Empirical evaluations indicate improved dialog realism and empathic accuracy, guiding future AI integration in therapeutic support tools.

Person-Centered Therapy (PCT) is a therapeutic modality rooted in humanistic psychology, with a primary emphasis on the client’s subjective experience and the creation of a facilitative relational climate. Central to the PCT paradigm are the conditions of empathic understanding, unconditional positive regard, and congruence. Recent computational advances, including the development of culturally attuned datasets and evaluation frameworks, have enabled operationalization and empirical benchmarking of PCT principles in LLMs and vision-LLMs (VLMs), with rigorous adherence to theoretical constructs and emerging empirical verification in diverse linguistic and cultural contexts (Abbasi et al., 9 Feb 2025, Tahir et al., 2 Sep 2025).

1. Theoretical Foundations of Person-Centered Therapy

Person-Centered Therapy was developed by Carl Rogers, who specified three necessary and sufficient therapeutic conditions: empathic understanding, unconditional positive regard, and congruence.

• Empathic Understanding is conceptualized as “the ability to perceive the client’s internal frame of reference as if it were your own, without ever losing the ‘as if’” (Rogers, 1957, 1961). Operationally, this involves therapist sensitivity to both expressed and unspoken emotions.
• Unconditional Positive Regard denotes non-judgmental acceptance of the client irrespective of their disclosures, strictly prohibiting evaluative or directive statements.
• Congruence (often referenced as genuineness) refers to transparent alignment between the therapist’s internal feelings and their outward communication. PCT also places high value on non-directive reflection and open inquiry, eschewing unsolicited advice in favor of client-led exploration (Tahir et al., 2 Sep 2025).

Recent computational research extends congruence to include multimodal emotional alignment, specifically investigating verbal-visual incongruence as diagnostic of underlying, unexpressed emotional states.

2. Computational Operationalization of PCT Principles

Datasets such as HamRaz and E-THER implement PCT’s facilitative conditions through structured prompt engineering, multimodal annotation, and evaluation schemas.

• HamRaz builds a hybrid architecture combining script-based templates and LLM-driven role-play to generate realistic, culturally attuned Persian dialogues spanning five canonical PCT stages: rapport building, active/empathetic listening, self-exploration, growth/change, and closing (Abbasi et al., 9 Feb 2025).
  • Empathy is enforced via prompts that require reflection and explicit labeling of client affect.
  • Unconditional positive regard is encoded as a constraint prohibiting judgment and requiring validation markers in every response.
  • Congruence is prompted by encouraging therapist self-disclosure when clinically appropriate.
• E-THER operationalizes empathic understanding by training VLMs to detect incongruence between verbal statements and nonverbal cues, using multidimensional annotation (incongruence, engagement, valence, arousal, dominance) on synchronized video-utterance pairs. The dataset penalizes prescriptive language and rewards open, non-directive reflection (Tahir et al., 2 Sep 2025).

3. PCT-Grounded Evaluation Frameworks

Rigorous evaluation of PCT-aligned systems requires both generative dialogue metrics and adherence to therapeutic process scales.

HamRazEval Metrics

• General Dialogue Metrics: Six attributes (coherence, engagement, fluency, diversity, humanness, balance) rated on a 1–10 scale, summarized by

$$\text{GeneralScore} = \frac{1}{6}\left(\mathrm{Coherence} + \mathrm{Engagement} + \mathrm{Fluency} + \mathrm{Diversity} + \mathrm{Humanness} + \mathrm{Balance}\right)$$

• Therapeutic Relationship (mini-BLRI): A 12-item Barrett-Lennard Relationship Inventory, grouped into Empathy, Unconditional Positive Regard, and Congruence subscales:

$$\begin{aligned} \mathrm{Empathy} &= \sum_{i\in\{2,6,8,11\}} s_i \ \mathrm{Uncond.\,Pos.\,Regard} &= \sum_{i\in\{1,3,5,7,12\}} s_i \ \mathrm{Congruence} &= \sum_{i\in\{4,9,10\}} s_i \ \mathrm{BLRI} &= \mathrm{Empathy} + \mathrm{Uncond.\,Pos.\,Regard} + \mathrm{Congruence} \end{aligned}$$

E-THER Metrics

• Empathic Authenticity (EA): Proportion of responses containing genuine empathic markers, eschewing clichés.
• Responsive Engagement (RE): Proportion of responses situationally tailored to the client’s unique verbal and nonverbal communication.
• Therapeutic Concision (TC): Proportion of responses using concise, non-directive language.
• Composite PCT Adherence (PA): Arithmetic mean of EA, RE, and TC.

4. Empirical Outcomes in PCT-Aligned AI Systems

Empirical studies leveraging HamRaz and E-THER report statistically significant improvements in both therapeutic process and conversational realism.

Method	BLRI	GeneralScore
Script Mode	1.84	8.06
Two-Agent	1.58	8.03
HamRaz	2.85	9.31–9.55

LLM models fine-tuned on HamRaz achieve ΔBLRI ≈ +1.0–1.3 and ΔGeneral ≈ +1.2–1.5 over Script Mode and Two-Agent baselines ($p<0.01$). These differences reflect substantially increased demonstration of empathic understanding, unconditional positive regard, and congruence in generated dialogues (Abbasi et al., 9 Feb 2025).

For vision-LLMs trained on E-THER, VideoLLAVA displays Empathic Authenticity of 91.8% (vs 72.0% for GPT-4V, $p<0.001$), and Therapeutic Concision of 64.8% (vs 57.0%, $p<0.001$). Responsive Engagement improvements are observed but not statistically significant, underscoring the unique impact of incongruence-aware training (Tahir et al., 2 Sep 2025).

5. Multimodal Extensions: Verbal-Visual Incongruence and Engagement Dynamics

E-THER introduces systematic annotation and quantification of verbal-visual incongruence, where emotional valence and intensity of speech do not match facial or behavioral cues. Three incongruence types are distinguished: “none” (congruent), “minimizing” (visual distress exceeds verbal), and “contradiction” (visual affect opposes verbal content). The engagement metric ($E_i \in [0.0, 1.0]$) is evaluated per utterance, accommodating dynamic modeling of therapeutic alliance.

A formal incongruence score is given by:

$$s_i = \mathrm{clip}\left( \frac{ \| e_i^{(v)} - e_i^{(t)} \|_2 }{ \tau_e } + \lambda (1 - \langle z_i^{(v)}, z_i^{(t)} \rangle ),\ 0, 1 \right)$$

where $e_i^{(v)}, e_i^{(t)}$ are VAD (valence-arousal-dominance) vectors from visual/text cue channels and $z_i^{(v)}, z_i^{(t)}$ are normalized cross-modal embeddings. This operationalization is directly linked to PCT’s theoretical construct of congruence and provides a training signal for AI to detect unspoken distress, thereby minimizing “performative empathy” (Tahir et al., 2 Sep 2025).

6. Cultural and Linguistic Adaptation

PCT datasets targeting under-represented linguistic groups face unique challenges, including the absence of resources reflecting specific societal norms, idioms, and culturally-coded expressions. HamRaz addresses these by sourcing region-specific forums, injecting complexity via indirect and ambiguous statement prompts, and incorporating colloquial metaphors salient in Farsi (e.g., “بار سنگین” for emotional burden). This approach demonstrates that authentic therapy simulation requires both granular linguistic adaptation and structural embedding of humanistic principles (Abbasi et al., 9 Feb 2025).

E-THER’s annotation schema is compatible with extensions to other languages and modalities, suggesting scalability for broader cross-cultural application. A plausible implication is that continued development of culture- and modality-aware PCT datasets may facilitate robust, generalizable empathic AI.

7. Implications and Future Directions

Recent computational advances in person-centered therapy simulation have introduced high-fidelity datasets, evaluation protocols, and empirical evidence for improved empathic attunement and dialogic realism in both LLMs and VLMs. This establishes a benchmark for non-clinical empathic assistants and psychotherapy support tools that operationalize humanistic psychotherapy principles.

Identified future directions include (a) scaling annotation using semi-automatic methods, (b) cross-theoretical benchmarking against other modalities (e.g., CBT, DBT), (c) multidimensional emotional reasoning via additional modalities (e.g., gesture, prosody), and (d) clinical and ethical validation against established empathy scales and therapist judgment. The explicit quantification and implementation of PCT adherence suggest pathways to responsible AI deployment in mental health–adjacent domains (Tahir et al., 2 Sep 2025).