---
title: Chain-of-Empathy (CoE) Framework
url: https://www.emergentmind.com/topics/chain-of-empathy-coe
type: topic
---

# Chain-of-Empathy (CoE) Framework

Chain-of-Empathy (CoE) is a prompting framework for large language models (LLMs) that makes response generation proceed through explicit reasoning about a user’s emotional state before a reply is produced. Introduced as a psychotherapy-inspired analogue to Chain-of-Thought prompting, it replaces direct response generation with an intermediate interpretive stage: the model first infers what emotion is present and what situational or cognitive factors may have caused it, and then uses that inference to choose an empathetic communication strategy [2311.04915]. In this formulation, empathy is treated not as surface politeness but as cognitive appraisal of emotional context, with the aim of producing context-aware, therapeutically informed, and safer responses, especially in mental-health-like dialogue settings [2311.04915].

## 1. Definition and conceptual scope

CoE was introduced in the paper "Chain of Empathy: Enhancing Empathetic Response of Large Language Models Based on Psychotherapy Models" [2311.04915]. The framework is presented as an empathy-oriented counterpart to Chain-of-Thought: instead of reasoning over logical steps in mathematics or symbolic tasks, the model reasons over human emotional experience and the causes behind it. The central motivation is that standard prompting in supportive dialogue often yields replies that are polite but shallow, generic, and predominantly exploratory, without being well tailored to the user’s actual emotional context [2311.04915].

The framework is explicitly procedural rather than parametric. It is not introduced as a mathematical model with learned parameters, but as a prompting procedure that inserts a structured interpretive stage between input and response. The paper’s common reasoning structure has two steps: identify any word or phrase that represents the client’s emotion, and understand the individual or situational factors that may have led to that expression [2311.04915]. The inferred state is then used to generate a response aligned with a target empathetic strategy.

A core conceptual claim of CoE is that empathetic understanding requires more than recognizing that a person is upset. It requires understanding why the person is upset, what kind of emotional pattern is occurring, and what kind of response is appropriate. In this respect, CoE frames empathetic dialogue as a structured inference problem rather than a one-shot stylistic generation problem. This suggests a shift from generic supportive language toward context-sensitive appraisal and response selection.

## 2. Prompting procedure and psychotherapy-derived variants

The most distinctive feature of CoE is its use of psychotherapy models as prompting scaffolds. The paper instantiates four psychotherapy-informed prompt variants, each designed to induce a different reasoning style: CBT-CoE, inspired by Cognitive Behavioral Therapy; DBT-CoE, inspired by Dialectical Behavior Therapy; PCT-CoE, inspired by Person-Centered Therapy; and RT-CoE, inspired by Reality Therapy [2311.04915]. The paper explicitly notes that these are simplified, prompting-oriented abstractions rather than complete clinical models.

The prompt design is zero-shot. In the base condition, the model is asked to reply directly, without reasoning. In the CoE conditions, the model is explicitly instructed to perform the two-step reasoning about emotion and causation before producing the response. This is the key methodological difference: standard prompting asks for a reply directly, whereas CoE forces an intermediate interpretive step that changes how the model understands the user message before answering [2311.04915].

The four variants differ in their reasoning focus and intended empathetic style.

| Variant | Reasoning focus | Goal or induced style |
|---|---|---|
| CBT-CoE | cognitive reframing; tackling negative thought patterns | identify negative thought patterns and cognitive errors, then reinterpret them |
| DBT-CoE | emotion regulation; addressing emotional dysregulation | recognize intense affect, validate distress, and help stabilize emotion |
| PCT-CoE | self-understanding; enhancing self-awareness | acceptance, reflective understanding, and self-exploration |
| RT-CoE | problem-focused coping; identifying causes of dissatisfaction | connect the emotional problem to current choices, needs, and possible actions |

CBT-CoE is centered on cognitive reframing. Its reasoning focus is described as “tackling negative thought patterns,” and the paper’s example figure associates it with recognizing “cognitive error” [2311.04915]. In the example “I just broke up. My life is over,” a CBT-like interpretation treats “My life is over” as a cognitive catastrophic error, which then licenses a response combining validation with reframing. DBT-CoE, by contrast, is organized around “addressing emotional dysregulation” and therefore emphasizes validation, soothing, grounding, and present-moment stabilization. PCT-CoE emphasizes unconditional positive regard, acceptance, and reflective listening, tending toward empathetic exploration rather than early interpretation or advice. RT-CoE is more action-oriented, linking dissatisfaction to choices, needs, and possible coping actions [2311.04915].

A frequent misconception is to treat these prompt variants as clinical implementations of psychotherapy. The source paper rejects that reading: the variants are deliberately simplified abstractions of therapeutic traditions, adapted for prompting rather than for formal therapeutic practice. A related misconception is that CoE simply asks the model to be warmer. Its actual mechanism is the insertion of a structured emotional-causal reasoning stage prior to generation.

## 3. Evaluation framework and empirical findings

The original CoE paper evaluates the method using OpenAI GPT-3.5 via the text-davinci-003 API, with temperature 0.9, top-p 1, frequency penalty 0, and presence penalty 0.6 [2311.04915]. The evaluation uses the EPITOME dataset, a crowdsourced set of Reddit help-seeking posts in mental-health contexts paired with response posts. The task focuses on three empathy strategy classes from the EPITOME framework: emotional reaction, exploration, and interpretation. The communication-strength scale uses levels 0, 1, and 2, where level 0 means no empathy expression and is excluded from analysis [2311.04915].

The class distribution was balanced by randomly sampling 500 emotional reaction examples and 500 interpretation examples while keeping 480 exploration examples, yielding a final set of 500/480/500 [2311.04915]. For each seeker post, the model was asked to generate a response and then predict which empathy strategy was most appropriate. The predicted strategy was compared to the ground-truth EPITOME label, and performance was measured using accuracy, precision, recall, and F1. Outputs with malformed or out-of-set strategy names were excluded because generative models sometimes produced labels such as “No Empathy Strategy” or novel labels like “Reflection” or “Validation” [2311.04915].

The reported pattern is clear. Without reasoning, the model strongly preferred exploration and showed the weakest use of interpretation. With CoE prompting, the models still leaned toward exploration, but they produced more interpretation than the base condition. Among the psychotherapy variants, CBT-CoE was the most balanced, generating a comparatively broader mix of strategies and the highest number of interpretation responses among the CoE conditions [2311.04915].

The paper interprets this outcome through the structure of CBT itself. Because CBT-style prompting asks the model to identify and correct cognitive distortions, it naturally moves generation beyond generic supportive language toward meaning-based interpretation, while still preserving validation and practical empathy. This suggests that structured therapeutic decomposition can alter not only response tone but also the distribution of empathic mechanisms the model deploys.

## 4. Response characteristics, safety significance, and qualitative behavior

Qualitative examples in the CoE paper show that CoE outputs are generally courteous and comprehensive, often combining multiple empathy strategies in a single reply and frequently suggesting professional support [2311.04915]. A characteristic CoE response often begins with an interpretation of the user’s state, then proceeds to exploration or advice. The paper contrasts this with many human benchmark responses in the dataset, which can be brief or less structured.

A prominent example is a DBT-CoE response to a user distressed after witnessing parental violence. The generated reply validates fear, states that the user is not at fault, and asks how it can support them now; the paper notes that this differs from the original human response in the dataset, which was much less constructive [2311.04915]. This comparison is used to argue that explicit emotional-causal reasoning can produce more contextually organized support than direct generation alone.

The broader significance claimed for CoE is its role in safer and more effective human-AI communication in mental-health-like settings. The framework is intended to give the model a more principled way to choose whether to validate, explore, interpret, regulate emotion, or suggest action. The paper argues that inappropriate or overly shallow responses can be harmful or unhelpful, and that psychotherapy-informed reasoning offers a mechanism for more context-specific empathetic behavior [2311.04915].

This should not be read as evidence of clinical adequacy. The paper’s own framing is more limited: CoE is presented as a mechanism for shaping LLM behavior toward context-sensitive empathy, not as a substitute for psychotherapy or as a validated counseling system.

## 5. Related formulations, extensions, and adjacent frameworks

Subsequent and related work has broadened the design space around empathy-oriented reasoning, sometimes explicitly invoking Chain-of-Empathy and sometimes using conceptually similar staged decompositions. In "Cause-Aware Empathetic Response Generation via Chain-of-Thought Fine-Tuning" [2408.11599], empathetic generation is organized as a chain of emotion recognition, emotion-cause extraction, cause-oriented commonsense expansion, listener-aware stance, and response generation. That work argues that emotion labels alone are insufficient because they capture state but not trigger, and it reports that cause-aware prompting plus instruction tuning improves both automatic and human evaluations on EmpatheticDialogue [2408.11599].

In spoken dialogue, "Leveraging Chain of Thought towards Empathetic Spoken Dialogue without Corresponding Question-Answering Data" [2501.10937] presents the Listen, Perceive, and Express (LPE) framework, which decomposes empathetic spoken response generation into three explicit substeps: listen to speech content, perceive the speaker’s emotion, and then express an empathetic response. The paper states that the best prompt is a zero-shot CoT with explicitly defined steps—transcribe the speech, identify the emotion, and generate an empathetic response—and reports that more verbose reasoning degrades performance [2501.10937]. The progression from speech understanding to affect inference to empathetic utterance is not named CoE in that work, but it functions as a speech-grounded empathy chain.

A data-centric alternative appears in "SoulChat: Improving LLMs' Empathy, Listening, and Comfort Abilities through Fine-tuning with Multi-turn Empathy Conversations" [2311.00273]. Rather than inserting an explicit reasoning prompt, SoulChat trains a model on a multi-turn corpus in which empathetic behavior is distributed across conversation turns: listening, recognition, validation, comfort, trust-building, questioning, and only then more solution-oriented content. The paper describes this as a move away from advice-first behavior and reports higher expert-rated empathy than ChatGLM-6B, MeChat, and ChatGPT on both SoulChatCorpus and SMILECHAT [2311.00273]. This suggests that CoE-like behavior can be induced either through prompt-level reasoning or through multi-turn supervision over empathic dialogue trajectories.

A more explicit extension is "Empathy-R1: A Chain-of-Empathy and Reinforcement Learning Framework for Long-Form Mental Health Support" [2509.14851]. That work formalizes a four-layer CoE structure—L1 emotions and context, L2 causes and beliefs, L3 intent analysis, and L4 response strategy—and trains a Chinese long-form support model using supervised fine-tuning followed by Group Relative Policy Optimization. It reports a Win@1 rate of 44.30% on Empathy-QA and provides an ablation directly comparing the model with and without CoE [2509.14851]. Relative to the 2023 prompting framework, this later formulation moves CoE from zero-shot prompt design toward an explicit train-time reasoning scaffold with reinforcement learning.

Other adjacent work expands the scope of empathy reasoning beyond text dialogue. "Nano-EmoX: Unifying Multimodal Emotional Intelligence from Perception to Empathy" [2603.02123] frames affective modeling as a hierarchy of perception, understanding, and interaction, with a curriculum named P2E (Perception-to-Empathy). "Empathy Modeling in Active Inference Agents for Perspective-Taking and Alignment" [2602.20936] gives a computational account in which empathy is an explicit weighting parameter in social expected free energy, coupled with self-other model transformation and Bayesian opponent inference. These papers do not define the same prompting framework as the original CoE paper, but they support a broader view of empathy as staged inference from signals and context to socially aligned response or action.

The acronym “CoE” is also used in unrelated domains. "Video-CoE: Reinforcing Video Event Prediction via Chain of Events" [2603.14935] uses CoE to denote Chain of Events in video event prediction, and "CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification" [2503.15234] uses CoE to denote Chain-of-Explanation in vision interpretability. These usages are terminologically distinct from Chain-of-Empathy.

## 6. Limitations, controversies, and open research directions

The original CoE paper notes several limitations. First, evaluation is limited to task-specific strategy prediction rather than broader psychological empathy instruments such as the Interpersonal Reactivity Index. Second, the study measures model output quality but not user perception, even though people may respond differently to empathetic language from AIs than from humans. Third, the quantitative study uses only one model, GPT-3.5, and only one domain, mental health [2311.04915].

These limitations point to several methodological controversies. One is whether empathy should be evaluated through strategy classification at all. The CoE study operationalizes performance through EPITOME strategy labels and communication strength, which captures deployment of emotional reaction, exploration, and interpretation, but not necessarily perceived warmth, trust, or long-horizon therapeutic utility. Another is whether explicit reasoning traces are beneficial at inference time in all settings. The spoken-dialogue LPE work reports that adding extra reasoning text can hurt performance, implying that structure is useful but verbosity can become noise [2501.10937]. A plausible implication is that effective CoE-style prompting may require careful control over reasoning granularity.

A further issue concerns domain transfer. The CoE paper suggests extending the method to domains such as career coaching or motivational interviewing and recommends broader datasets, multiple LLMs, and user-centered evaluation [2311.04915]. Related work supports this trajectory: cause-aware CoT has been applied to EmpatheticDialogue [2408.11599], long-form Chinese support has been studied through SFT and RL [2509.14851], and multimodal affective interaction has been framed as a perception-to-empathy curriculum [2603.02123]. This suggests that Chain-of-Empathy is less a single fixed prompt than a family of designs organized around a common principle: empathetic generation improves when the model is required to infer emotional causes, contextual meaning, and response strategy before answering.

At the same time, the literature repeatedly warns against overinterpretation. The psychotherapy-informed variants in the original CoE paper are simplified abstractions rather than clinical systems [2311.04915]. SoulChat notes that empathy and helpfulness can conflict, and that models fine-tuned without RLHF may still produce unsafe outputs [2311.00273]. Empathy-R1 positions itself as a support tool rather than a replacement for clinicians [2509.14851]. The most defensible reading, therefore, is that CoE is a research framework for structured empathic reasoning in LLMs, not a claim that current systems possess human empathy or therapeutic competence.

Source: https://www.emergentmind.com/topics/chain-of-empathy-coe