Therapist Empathy Scale (TES)

• TES is a 9-item observer rating scale that measures therapist empathy through affective, cognitive, attitudinal, and attunement dimensions.
• The scale can be scored as a total sum (9 to 63) or as an averaged item mean, allowing flexible analysis in both clinical and computational studies.
• In computational research, TES guides representation learning and evaluates therapy-capable language models by measuring process-level empathy cues.

Therapist Empathy Scale (TES) is a 9-item observer rating scale for therapist empathy that has been used to assess affective, cognitive, attitudinal, and attunement aspects of therapeutic responding. In recent computational psychotherapy research, TES functions both as a direct annotation instrument for whole counseling sessions and as an evaluation target for therapy-capable LLMs. Across these uses, TES is treated as a measure of whether therapist behavior is therapeutically empathic rather than a measure of patient improvement or general conversational quality (Tao et al., 2023, Tahir, 8 Sep 2025).

1. Construct and conceptual scope

TES is defined in the counseling literature summarized in recent arXiv work as an observer-based instrument for therapist empathy. In the counseling setting, empathy is framed in Rogers-style terms as “the therapist’s sensitive ability and willingness to understand the client's thoughts, feelings, and struggles from the client's point of view.” The instrument comprises 9 items spanning affective, cognitive, attitudinal, and attunement dimensions, making it broader than a narrow measure of warmth or verbal affirmation alone (Tao et al., 2023).

In the LLM-based ACT study, the same scale is described as “a modality agnostic tool used to evaluate therapeutic empathy.” That characterization is important because it separates the construct from any particular delivery medium. TES is therefore not tied, in principle, to face-to-face human counseling, audio-only data, or text-only simulated sessions. What it evaluates is whether therapist responses exhibit empathic and therapeutically attuned qualities. In that study, TES is explicitly distinguished from patient improvement and from generic dialogue quality, and is used instead to assess whether the model’s therapist responses feel empathic in simulated therapy sessions (Tahir, 8 Sep 2025).

A common misconception is that therapist empathy scales merely track politeness or emotional positivity. The studies summarized here do not support that reduction. TES is treated as a structured observer rating of empathic therapeutic process, and in the ACT setting the authors further connect high TES to stance, pacing, reflective listening, and other process-level cues rather than to isolated phrases (Tahir, 8 Sep 2025).

2. Scoring conventions and unit of analysis

The TES format reported in the counseling representation-learning study is straightforward: each of the 9 items is scored from 1 to 7, so the total score ranges from 9 to 63, with higher TES indicating higher therapist empathy. In that setting, the rating is assigned to the entire counseling conversation, not to a local span of turns. The paper therefore treats empathy as a global subjective assessment of the session as a whole (Tao et al., 2023).

The ACT LLM study uses a different aggregation for analysis. There, the reported endpoint is the “TES total empathy mean,” defined as “the average of all nine TES items.” The study explicitly notes that the analytic score is not a total sum but the mean of the 9 rubric items, yielding a session-level empathy score for each generated transcript. Each transcript corresponds to a 25-turn interaction, and every transcript is rated on TES by an LLM judge (Tahir, 8 Sep 2025).

These two scoring conventions are compatible but analytically distinct. One preserves the original summed format with range 9 to 63; the other rescales the same rubric to an item mean. This suggests that TES can be deployed either as a conventional whole-session observer score or as a normalized session-level metric, depending on the modeling and statistical framework. What remains constant across both uses is the unit of substantive interpretation: TES is a rating of therapist empathy at the session level, not a direct measure of client outcome.

3. TES as supervision for representation learning from counseling conversations

A direct computational use of TES appears in "Learning Representation of Therapist Empathy in Counseling Conversation Using Siamese Hierarchical Attention Network" (Tao et al., 2023). The study uses the CUEMPATHY dataset, consisting of 156 audio recordings, 39 therapist-client dyads, and 4 conversations per dyad, with sessions of about 50 minutes each in Cantonese counseling practicums in a university setting. A trained observer rates each conversation using TES, and that session-level score becomes the target signal for learning conversation representations.

Because counseling sessions are long and contain many turns, the model organizes the dialogue hierarchically. Speaker turns are grouped into sections, and sections compose the full conversation. The encoder is a Siamese hierarchical attention network using bidirectional GRU (BiGRU) at both the turn and section levels. Turn-level attention is defined by

$u_{i,j} = \tanh(W_t h_{i,j} + b_t),$

$$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$

and the section embedding is

$s_i = \sum_j \alpha_{i,j} h_{i,j}.$

The same weighted-sum mechanism is then applied at the section level to obtain a final conversation embedding. The architectural motivation is explicit: TES is global, but empathic behavior may be concentrated in particular turns and local interaction spans.

The Siamese training objective uses pairs of conversations labeled by empathy similarity. If a pair comes from the same empathy group, $y=1$; if it comes from different empathy groups, $y=0$. The contrastive loss is

$$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$

where $d$ is the Euclidean distance between conversation embeddings and $m>0$ is a margin. The model is trained using only extreme TES examples, divided into low-empathy and high-empathy groups. For group size $K$, the procedure yields

$(2K^2 - K)$

pairs, including $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$0 positive pairs and $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$1 negative pairs. The 156 conversations are sorted by TES score, training uses the lowest $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$2 and highest $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$3 conversations, and testing uses the middle 96 conversations, $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$4, so that generalization is evaluated away from the extremes.

The reported findings show that the learned embedding space tracks TES structure. Distances between conversation embeddings are positively correlated with TES differences, and the embeddings can be used for Support Vector Regression (SVR) with leave-one-dyad-out cross-validation to predict TES on the held-out middle set. The best setting is $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$5, for which the mean absolute prediction error is $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$6 and 81 of the 96 test conversations have absolute error below 5.0. For $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$7, correlations include $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$8 to low-score references and $$\alpha_{i,j} = \frac{\exp(u_{i,j}^\top u_t)}{\sum_j \exp(u_{i,j}^\top u_t)},$$9 to high-score references. The paper also reports that a cross-entropy baseline does not predict TES well, whereas the contrastive-loss encoder performs substantially better. In methodological terms, TES here serves as the central supervision signal for learning a structured representation of therapist empathy from long-form counseling dialogue.

4. TES in evaluation of therapy-capable LLMs

In "The Thinking Therapist: Training LLMs to Deliver Acceptance and Commitment Therapy using Supervised Fine-Tuning and Odds Ratio Policy Optimization" (Tahir, 8 Sep 2025), TES is one of the two primary outcome measures used to assess whether a small open-weight LLM can produce therapeutically empathic responses during simulated ACT sessions. The study compares six therapist model conditions: Instruct (COT), Instruct (no COT), SFT (COT), SFT (no COT), ORPO (COT), and ORPO (no COT). Each model variant participates in 150 simulated therapy sessions, each session has 25 turns, and the same patient profiles are reused across model variants to control confounding. Chain-of-thought is handled internally: therapist models can generate thinking tokens after a <thinking> or Thinking: tag, but the simulated patient receives only the text after the Answer: prefix.

TES ratings are produced by an LLM judge: Mistral-Large, fine-tuned on human evaluated synthetic ACT transcripts. The paper does not provide the full TES rubric, item anchors, or exact judge prompt, and it does not reproduce the 9-item checklist verbatim. It does, however, specify that TES is used as an observer-style rating of therapist empathy and that the final metric is the mean of all nine items.

The primary mixed-effects analysis finds a statistically significant effect of model type on TES, with Wald $s_i = \sum_j \alpha_{i,j} h_{i,j}.$0. For TES, the paper reports marginal $s_i = \sum_j \alpha_{i,j} h_{i,j}.$1, conditional $s_i = \sum_j \alpha_{i,j} h_{i,j}.$2, and a meaningful random intercept for patient profile, ICC = 0.18, 95\% CI [0.14, 0.22]. In the table, the random effect is reported as Patient Profile ID: 0.18 (17.8%), indicating that patient characteristics substantially influence TES scores.

Model condition	Mean TES (SD)
Instruct (COT)	5.29 (0.95)
Instruct (no COT)	5.45 (0.90)
SFT (COT)	5.07 (1.03)
SFT (no COT)	4.72 (1.34)
ORPO (COT)	5.68 (0.88)
ORPO (no COT)	5.76 (0.82)

Using Instruct (COT) as the reference condition with EMM = 5.29, the reported TES contrasts are: Instruct (no COT), $s_i = \sum_j \alpha_{i,j} h_{i,j}.$3 $s_i = \sum_j \alpha_{i,j} h_{i,j}.$4, Hedges $s_i = \sum_j \alpha_{i,j} h_{i,j}.$5, adjusted $s_i = \sum_j \alpha_{i,j} h_{i,j}.$6; ORPO (COT), $s_i = \sum_j \alpha_{i,j} h_{i,j}.$7 $s_i = \sum_j \alpha_{i,j} h_{i,j}.$8, $s_i = \sum_j \alpha_{i,j} h_{i,j}.$9, adjusted $y=1$0; ORPO (no COT), $y=1$1 $y=1$2, $y=1$3, adjusted $y=1$4; SFT (COT), $y=1$5 $y=1$6, $y=1$7, adjusted $y=1$8; and SFT (no COT), $y=1$9 $y=0$0, $y=0$1, adjusted $y=0$2. The narrative summary states that ORPO variants had the highest TES scores, that ORPO without COT was the best overall on TES, and that the difference between the two ORPO variants was not significant with mean difference $y=0$3 and adjusted $y=0$4. By contrast, adding COT to SFT significantly improved empathy with mean difference 0.35 and Holm–Sidak adjusted $y=0$5, while adding COT to Instruct did not significantly change empathy.

The study links these outcomes to training paradigm. The authors argue that ORPO improves TES because the odds-ratio objective increases the log-odds of the preferred continuation over a matched alternative, pushing the policy toward process-level cues such as stance, pacing, reflective listening rather than mere reproduction of surface content. They further argue that, for SFT, COT likely functions as a scaffold that compels deliberation about values, cognitive processes, and therapeutic stance before answer generation. In their interpretation, explicit reasoning is therefore conditionally useful: helpful for imitation-only training, but not beneficial for the stronger ORPO or instruct-tuned variants.

5. TES and adjacent automated empathy measures

Not all computational psychotherapy studies operationalize therapist empathy through TES. "Empirical Modeling of Therapist-Client Dynamics in Psychotherapy Using LLM-Based Assessments" does not use the classic TES instrument; instead, it constructs an LLM-based proxy measure grounded in EPITOME plus a separate Reflection indicator (Chen et al., 12 Feb 2026). That distinction is methodologically important because the resulting measure is TES-like only in a loose sense.

In that study, therapist empathy is decomposed into Emotional Reactions, Interpretations, Explorations, and Reflection. The first three components are drawn from EPITOME and each scored on a 3-point ordinal scale from 0 to 2, yielding a summed empathy score from 0 to 6; Reflection is modeled separately as binary. Final automated scoring uses GPT-4o-mini, with 3 completions averaged to reduce sampling variance, and the model evaluates therapist empathy at the utterance level using the therapist utterance plus 2 prior client utterances as context. Validation on a stratified sample uses 3 human raters per utterance. Reported LLM–human agreement includes ICC(2,k) = 0.742 and $y=0$6 for Emotional Reactions, ICC(2,k) = 0.757 and $y=0$7 for Interpretation, ICC(2,k) = 0.730 and $y=0$8 for Exploration, and ICC(2,k) = 0.658 and $y=0$9 for Reflection. The abstract summarizes the broader alignment across constructs as mean Pearson $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$0.

The paper explicitly notes that the references include “Development of the Therapist Empathy Scale (Decker et al., 2014)”, but that TES is not operationally administered. Instead, empathy is modeled through proxy dimensions and later reduced by PCA into General Empathy and Exploration, with a two-component solution explaining about 75% of variance for $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$1 therapist-behavior observations. This provides a useful contrast with TES-based work. TES supplies a direct, established observer scale for session-level empathy; the EPITOME-plus-Reflection framework supplies a text-only, fine-grained, turn-local proxy suitable for large-scale structural modeling. These should not be treated as interchangeable without qualification.

6. Methodological implications, correlations, and limitations

Across the recent literature, TES occupies a distinctive position because it allows empathy to be evaluated separately from broader therapy fidelity while still revealing strong relations to therapeutic process. In the ACT LLM study, TES correlates strongly with ACT fidelity subscales: $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$2 with consistent stance, $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$3 with consistent awareness, $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$4 with consistent openness, and $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$5 values from $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$6 to $$L = \frac{1}{2} y d^2 + \frac{1}{2}(1-y)\{\max(0, m-d)\}^2,$$7 for the inconsistency subscales (Tahir, 8 Sep 2025). The paper’s broader message is therefore that empathy and fidelity are separable constructs, but in that dataset they co-occur strongly.

The same study also shows that context matters. The random intercept for patient profile, reported as ICC = 0.18 with 95% CI [0.14, 0.22], indicates that patient characteristics materially influence TES scores. A plausible implication is that empathy assessment in therapy-capable LLMs should not be interpreted as a pure model property independent of interactional context. The therapist response, the patient profile, and the session trajectory jointly shape the observed TES outcome.

At the same time, automated TES-style evaluation has clear limitations. In the ACT work, the judge is Mistral-Large fine-tuned on human-evaluated synthetic ACT transcripts, but the study does not specify how many human-rated examples were used, whether the human ratings were ordinal or scalar, whether the judge was trained jointly on ACT-FM and TES, or whether TES-specific ratings were part of the fine-tuning target. The paper also acknowledges that an LLM judge cannot fully replace human clinical raters. The psychotherapy dynamics paper generalizes the caution: observer-rated constructs may not match client perceptions, text-only assessment misses audio and other multimodal cues, and culturally normative communication styles may be misread by LLMs (Chen et al., 12 Feb 2026).

Taken together, these findings position TES as a rigorous observer-based benchmark for therapeutic empathy that is compatible with both traditional counseling research and modern computational modeling. In session-level representation learning, it anchors a contrastive objective over long counseling conversations. In therapy-capable LLM evaluation, it supplies a direct criterion for whether generated responses are empathic and therapeutically attuned. In adjacent large-scale automated process modeling, TES also serves as a reference point against which proxy measures can be distinguished. The central methodological lesson is that empathy should be measured explicitly, not collapsed into fidelity, fluency, or rapport, even when those constructs are strongly associated in practice.