Using Linguistic Synchrony to Evaluate Large Language Models for Cognitive Behavioral Therapy
Synchrony describes responsive communication between individuals and is known to be important in building social relationships and supporting mental health outcomes (Delaherche et al., 2012; Klein, 2023). The phenomenon manifests through various modalities, including physical body movements (mirrored body language) (Ramseyer and Tschacher, 2011), vocals (pitch matching) (Imel et al., 2014), and language (linguistic style matching) (Niederhoffer and Pennebaker, 2002), across a variety of contexts (Kidby et al., 2023; Bonny and Jones, 2023). Synchrony is associated with building a sense of affiliation and improving cooperation and rapport (Vail et al., 2022); it is critical in therapist-client relationships (Colton, 2022). In this work, we focus on linguistic synchrony in the context of mental health therapy.
which we operationalize through the normalized Conversational Linguistic Distance (nCLiD) (Nasir et al., 2019), and two measures of the quality of client self-disclosures - intimacy and engagement. Then we compare the performance of the LLM to trained therapists and non-expert online peer supporters in a CBT setting (Figure 1). We show that the LLM is outperformed by both groups. This indicates that LLMs are not yet at the level of humans in generating high-quality therapeutic responses, and we suggest that synchrony can serve as an evaluation criterion for LLMs in mental health contexts
Descriptive intimacy involves the dis closure of private facts, while evaluative intimacy involves the disclosure of personal opinions and information. Engagement is the extent to which a patient actively participates in the therapeutic process beyond simply being present
The concept of interpersonal synchrony describes when the participants of an interaction adapt and converge on each other’s behaviors over time.
focusing on linguistic synchrony, which refers to the similarity between interlocutors in semantics, syntax, or style.
Both Cho et al. (2023) and (Chiu et al., 2024) simulate the client side of the LLM-client conversation due to ethical concerns of having an LLM advise vulnerable populations. However, this prevents a realistic evaluation of LLMs for therapy. The LLM-participant dataset used in this work comes from an IRB-approved study (Kian et al., 2024) in which they deployed LLMs in an interactive CBT homework context with students (Section 4), which provides a step towards more realistic evaluations of LLMs in therapy
framework. Descriptive intimacy involves the disclosure of private facts, while evaluative intimacy involves the disclosure of personal opinions and feelings.
The CBT exercise transcripts were annotated for three variables: descriptive intimacy, evaluative intimacy, and engagement. Four undergraduate annotators (two female, two male) were trained through workshops led by graduate student instructors for two weeks to annotate the data for the selected variables.
This means that higher synchrony is associated with higher intimacy and active engagement (Supporting H1a, H1b, and H1c). We hypothesize that in a therapeutic setting, a therapist’s linguistic synchrony with the client encourages greater self-disclosures and, subsequently, higher levels of intimacy and engagement.
Colton (2022) also found that linguistic synchrony “catalyzes” the therapeutic bond, which is further supported by Vail et al. (2022).
The significant relationship between linguistic synchrony and descriptive intimacy, evaluative intimacy, and engagement indicates that nCLiD shows promise as a measure of therapeutic outcomes.
Lines of inquiry this paper opens 24
Research framings built by reading the notes related to this paper — the questions it feeds into.
Can real-time computational alliance measurement improve therapy outcomes?- Why do therapists and patients report misaligned perceptions of the working relationship?
- Why does therapist 'we' language also predict lower therapeutic alliance?
- Why might patients feel closest to therapists when misalignment is highest?
- Can working alliance be measured in real time during therapy sessions?
- Does therapist alliance perception function like expressed satisfaction rather than actual progress?
- Which therapy topics increase alliance scores across different mental health conditions?
- Can therapists use real-time alliance scores to adjust their approach during sessions?
- Can single-turn empathy scores predict performance in ongoing therapeutic relationships?
- Can single-turn empathy advantage predict multi-turn therapeutic outcomes?
- How does linguistic synchrony differ between LLMs and human therapists over time?
- How do patient filler pauses signal safety and trust in therapy?
- Can real-time pronoun feedback improve therapist training outcomes?
- How does linguistic synchrony between therapist and client predict disclosure?
- How do trained therapists and peer supporters differ from LLMs on conversational synchrony?
- Does true understanding matter for therapeutic benefits of disclosure?
- How much does impression management prevent honest self-disclosure?
- How does self-disclosure function as a common ground building act?
- What is the relationship between pronoun patterns and linguistic entrainment?
- Does linguistic coordination signal both therapeutic rapport and manipulative intent?
- How does lexical entrainment differ between human therapists and conversational AI?
- Does linguistic coordination always signal rapport or can it indicate manipulation instead?
- What does it mean to truly attend to someone in conversation?