EmoTrace: An Emotion Trajectory-Centered Framework for Psychological Support Dialogue Generation

# EmoTrace: An Emotion Trajectory-Centered Framework for Psychological Support Dialogue Generation
Source: [https://arxiv.org/html/2607.23648](https://arxiv.org/html/2607.23648)
Kaitong Weng, Lixin Liu, Zihao Liu, Bo Wang111Corresponding author\., Shiguang NI111Corresponding author\. Shenzhen International Graduate School, Tsinghua University \{wengkt25, liulx25, liuzh25\}@mails\.tsinghua\.edu\.cn bo\-wang@tsinghua\.edu\.cn, ni\.shiguang@sz\.tsinghua\.edu\.cn

###### Abstract

Using large language models \(LLMs\) to assist psychological counseling is an important task in the field of natural language processing\. The construction of high\-quality psychological support dialogue corpora serves as a critical foundation for training counseling\-oriented conversational models\. However, existing data generation approaches generally suffer from several limitations, including emotionally stable seekers, limited variation in emotional dynamics, and a high degree of compliance with counselors’ guidance\. These issues result in LLM that lack the capability to effectively respond to emotionally unstable scenarios\. In addition, counselor responses are typically driven by problem\-solving objectives, thereby overlooking the role of emotion\-focused interaction, which are essential in psychological counseling\. To address these gaps, we proposeEmoTrace, a multi\-turn dialogue corpus generation framework centered on modeling seekers’ emotional trajectories\. we construct seekers’ cognitive profile and introduce a seeker module with emotional schemas and an associated activation mechanism, a counselor module, and an emotional trajectory control module, thereby enhancing the layering of the seeker’s emotional expression and the counselor’s targeted empathic expression\. Experimental results demonstrate that the proposed method outperforms existing approaches in terms of emotional richness and empathy quality\. The complete dataset and the fine\-tuned model will be made publicly available once the paper is accepted\.

EmoTrace: An Emotion Trajectory\-Centered Framework for Psychological Support Dialogue Generation

Kaitong Weng, Lixin Liu, Zihao Liu, Bo Wang111Corresponding author\., Shiguang NI111Corresponding author\.Shenzhen International Graduate School, Tsinghua University\{wengkt25, liulx25, liuzh25\}@mails\.tsinghua\.edu\.cnbo\-wang@tsinghua\.edu\.cn, ni\.shiguang@sz\.tsinghua\.edu\.cn

## 1Introduction

Mental health constitutes a fundamental component of overall health, with approximately 14% of the global disease burden attributable to neuropsychiatric disorders\(Princeet al\.,[2007](https://arxiv.org/html/2607.23648#bib.bib1)\)\. However, due to its longstanding underrepresentation in global health priorities, mental health services continue to face stigma and limited accessibility\. As a result, many affected individuals are unable to obtain timely care\(VanderWeeleet al\.,[2026](https://arxiv.org/html/2607.23648#bib.bib2)\)\.

The rapid development of large language models \(LLMs\) has introduced new perspectives for psychological counseling applications\. Models such as GPT\(Achiamet al\.,[2023](https://arxiv.org/html/2607.23648#bib.bib3)\)and LLaMA\(Touvronet al\.,[2023](https://arxiv.org/html/2607.23648#bib.bib4)\)have demonstrated human\-like emotional support capabilities\(Sorinet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib37)\)\. Building on this progress, a number of studies have proposed specialized models for the counseling domain, including MeChat\(Qiuet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib5)\)and CpsyCounX\(Zhanget al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib6)\)\. These models rely on real\-world or synthetic dialogue corpora to learn general patterns of psychological counseling\. Consequently, the construction of high\-quality dialogue corpora becomes a critical factor influencing model performance\(Qiet al\.,[2025](https://arxiv.org/html/2607.23648#bib.bib18)\)\. However, due to constraints related to medical ethics and personal privacy protection, the construction of large\-scale, compliant mental health support dialogue datasets remain a central challenge in developing LLM\-based counseling systems\(Mayeret al\.,[2022](https://arxiv.org/html/2607.23648#bib.bib13)\)\.

![Refer to caption](https://arxiv.org/html/2607.23648v1/x1.png)Figure 1:A comparison of the emotion trajectories of seekers in traditional datasets and EmoTrace\-D on the topic of social phobia\.Recently, several studies on automated corpus construction, such as DeepWell\-Adol\(Qiuet al\.,[2025](https://arxiv.org/html/2607.23648#bib.bib8)\)and PsyDT\(Xieet al\.,[2025](https://arxiv.org/html/2607.23648#bib.bib9)\), have focused on modeling and optimizing the counselor side of the psychological support dialogue\. These approaches improve response quality by simulating counselors’ linguistic styles and therapeutic orientations or by employing dialogue frameworks grounded in positive psychology\.

However, this counselor\-centered paradigm for corpus construction exhibits some limitations\. As shown in[Figure 1](https://arxiv.org/html/2607.23648#S1.F1), the seekers’ emotional states in conventional dialogue often show low variability and a homogeneous progression, accompanied by a high level of compliance\. Such patterns fail to reflect the emotional fluctuations commonly observed in real\-world interactions\. Training corpora constructed by this paradigm make it difficult for models to learn response strategies in the face of complex emotional dynamics\. Furthermore, existing work typically characterizes seekers using emotion labels\(Liuet al\.,[2021](https://arxiv.org/html/2607.23648#bib.bib10)\)and the Big Five personality traits\(Yanget al\.,[2025b](https://arxiv.org/html/2607.23648#bib.bib11)\)\. However, they hardly considered the underlying cognitive structures of emotions, such ascore beliefs\(individuals’ deepest self\-perceptions and value systems\)\(Becket al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib16)\)andemotional schemas\(cognitive evaluation and coping patterns of emotional experiences\)\(Leahy,[2002](https://arxiv.org/html/2607.23648#bib.bib17)\)\. This approach limits seekers to narrating only events, lacking cognitive evaluations of core beliefs or emotional schemas, so the model struggles to capture the logical chain from event to cognition to emotion\.

Additionally, prior corpus construction work also have certain limitations in the design of counselor response strategies\. Within the humanistic therapy tradition, Emotion\-Focused Therapy \(EFT\) represents a quintessential approach\(Greenberg,[2004](https://arxiv.org/html/2607.23648#bib.bib12)\)\. EFT emphasizes the acceptance and integration of emotional experience, adhering to a seeker\-centered framework\. However, in existing work, counselor responses are largely oriented toward direct problem\-solving, while emotional processing is confined to surface\-level empathy\. This approach neglects the central role of emotion in the process of psychological support\(Greenberg,[2010](https://arxiv.org/html/2607.23648#bib.bib14)\)\.

![Refer to caption](https://arxiv.org/html/2607.23648v1/x2.png)Figure 2:The overall framework of EmoTrace\.In this paper, we proposeEmoTrace, a seekerEmotionTrajectory\-Centered framework for multi\-turn psychological support dialogue generation\. Our method employs interactive role\-playing to simulate dialogues between a seeker and a counselor\. Specifically, in the seeker profile construction stage, we add basic personal information with an emotional schema dimension\. The corpus generation is accomplished through the collaboration of three core modules\. The overall framework of EmoTrace is illustrated in[Figure 2](https://arxiv.org/html/2607.23648#S1.F2)\.The seeker modulegenerates responses that maintain personality consistency by integrating the persona profile with an emotional schema activation mechanism, while applying frequency constraints to avoid excessive schema activation\.The counselor module, grounded in EFT, produces more targeted intervention strategies and responses\. Finally,the Emotion Trajectory \(ETC\) Control moduleguides the seeker’s emotional development using a three\-stage structure, modeling emotion as a trend variable to achieve controllable trajectories without compromising naturalness\.

Based on this method, we constructEmoTrace\-D, a psychological support dataset containing 1,114 multi\-turn dialogues\. We also fine\-tune an open\-source LLM namedEmoTrace\-Mon EmoTrace\-D\. Extensive experimental results demonstrate that, compared with prior work, the proposed method significantly improves the emotional reality and complexity of the generated corpora and boosts the empathic alignment of the model when using the data\. Our contributions are as follows:

- •We proposeEmoTrace, the first paradigm shifting corpus construction from counselor\-centered to seeker emotional trajectory\-centered, enabling seekers as dynamic emotional entities for more authentic dialogues\.
- •We construct theEmoTrace\-Dand fine\-tune theEmoTrace\-Mon this dataset\. Extensive empirical experiments demonstrate that the model achieves superior targeted empathic capability compared to prior work\.

## 2Related Work

### 2\.1Psychological Support Dialogue Datasets

Early psychological support dialogue datasets primarily relied on crowdsourcing and online psychological counseling platforms as data sources, with support strategies annotated manually\(Welivita and Pu,[2022](https://arxiv.org/html/2607.23648#bib.bib25); Sunet al\.,[2021](https://arxiv.org/html/2607.23648#bib.bib26); Wanget al\.,[2019](https://arxiv.org/html/2607.23648#bib.bib27)\)\. However, due to the highly sensitive nature of psychological counseling, the acquisition of relevant corpora is often subject to privacy restrictions\. Consequently, the majority of existing research has adopted synthetic data approaches to construct corpora\. For instance, SmileChat\(Qiuet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib5)\)extends PsyQA from single\-turn question\-answering to multi\-turn dialogues using ChatGPT, while CpsyCounD\(Zhanget al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib6)\)employs a two\-stage framework based on counseling reports to generate 3,134 multi\-turn dialogues\.

Recently, to enhance the professionalism of psychological support dialogue corpora, some studies integrate psychological theories into the dialogue generation process\(Leeet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib28); Zhouet al\.,[2025b](https://arxiv.org/html/2607.23648#bib.bib29),[a](https://arxiv.org/html/2607.23648#bib.bib31); Shiet al\.,[2026](https://arxiv.org/html/2607.23648#bib.bib30)\)\. However, these methods primarily focus on the selection and optimization of counseling strategies, without the modeling of the seeker’s cognitive structure\.

### 2\.2Psychological Counseling Dialogue Systems

The rapid development of LLMs has advanced the growth of AI\-driven psychological counseling services\. Recent studies, such as MindChat\(Xin Yan,[2023](https://arxiv.org/html/2607.23648#bib.bib7)\)and SoulChat\(Chenet al\.,[2023](https://arxiv.org/html/2607.23648#bib.bib32)\), apply supervised fine\-tuning to foundation models using high\-quality dialogue data, with a focus on improving empathetic capabilities\.

Given that psychological counseling involves tasks such as intervention strategy planning and empathetic response generation, it naturally aligns with the collaborative paradigm of multi\-agent systems, leading to the growing adoption of multi\-agent approaches for psychological counseling\. For example, WiseMind\(Wuet al\.,[2026](https://arxiv.org/html/2607.23648#bib.bib33)\)employs two cooperative agents representing rational reasoning and emotional reasoning, together with a DSM\-5 knowledge graph for psychiatric assessment, thereby improving diagnostic accuracy and reducing hallucinations\. To address difficulties in cross\-session memory retrieval, TheraMind\(Huet al\.,[2026](https://arxiv.org/html/2607.23648#bib.bib34)\)adopts a dual\-loop agent architecture designed for long\-term counseling, with capabilities for strategy planning and adaptive adjustment\. However, these methods lack reasoning about the causes of the seeker’s emotional changes, making it difficult to handle complex emotional scenarios\.

## 3Methodology

### 3\.1Overview of the Psychological Support Dialogue Generation Framework

We synthesize psychological support dialogues through interactive role\-playing and core component control, including three modules: a seeker module, a counselor module, and an ETC module\.

Firstly, based on the persona profile, the seeker module generates responses by integrating the emotional expression guidance provided by the ETC module with the activation status of emotional schemas\. Then, the counselor module generates responses via a pipeline consisting of two submodules: the analysis\-planning submodule formulates psychological intervention strategies by emotion analysis and schema recognition, while the generation submodule then produces responses based on the intervention strategy derived from the former\. After each turn of interaction between the seeker and the counselor, the ETC module determines the seeker’s current stage based on the dialogue history\. It then infers the emotional dynamics for the next turn, and provides guidance for subsequent emotional expression\.

### 3\.2Design of Psychological Support Dialogue Generation

#### 3\.2\.1Preparation: Seeker Persona Profile Construction

To enhance the authenticity and consistency of the seeker’s expression patterns in psychological support dialogue corpora, we draw inspiration from personalized role\-driven study\(Wanget al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib15)\), constructing persona profiles by extracting individual characteristics and psychological attributes from raw dialogue data\. This enables the LLM\-driven seeker to maintain consistent cognitive patterns and emotional response styles across multi\-turn dialogues\. We used 5,000 multi\-turn dialogues from PsyDT as seed data and employed GPT\-4\.1\-mini to summarize the seeker’s background information and personality traits in each dialogue, organizing them into structured persona profiles\. Each profile contains information such as gender, age, occupation, interaction style and problems\.

Traditional persona profile construction typically characterizes individuals merely with emotion labels or simple background information, which fails to capture their underlying psychological mechanisms\. To address this limitation, we propose the emotional schema as a modeling dimension to characterize individual traits across three levels: cognitive, emotional, and behavioral\. Specifically, based on the classification of core beliefs in Cognitive Behavioral Therapy \(CBT\)\(Becket al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib16)\)and the emotional schema model\(Leahy,[2002](https://arxiv.org/html/2607.23648#bib.bib17)\), we establish a three\-dimensional cognitive axis:the self axis\(cognitive style of one’s own emotions\),the others axis\(cognitive style of others and interpersonal relationships\), andthe world/future axis\(cognitive style of the persistence and meaning of emotions\)\. Each axis reflects how individuals respond to emotional experiences from different perspectives\. On this basis, we select eight representative emotional schemas, each linked to typical cognitive tendencies along these axes\. For example, people with a guilt schema tend to feel self\-reproach when their own actions negatively affect others\. Detailed information on emotional schemas is provided in the Appendix[A\.1](https://arxiv.org/html/2607.23648#A1.SS1)\.

Finally, we clustered and filtered all persona profiles according to topics and schema types, resulting in 1,423 uniformly distributed persona profiles\. The prompt for generating persona profiles is provided in the Appendix[A\.1](https://arxiv.org/html/2607.23648#A1.SS1)\.

#### 3\.2\.2Emotional Trajectory Control Module

This module provides controllable guidance of the seeker’s emotional development process across multi\-turn dialogues based on three\-stage emotional trajectory modeling\.

Stage\-based emotional trajectory\.Emotional change in psychological counseling is gradual, not instantaneous\(Stileset al\.,[1990](https://arxiv.org/html/2607.23648#bib.bib35)\)\. Therefore, observing emotional change from a trajectory perspective can better captures the continuity of the seeker’s emotional state and the process of emotional evolution than single time point judgments\.

Inspired by the EFT theory\(Greenberg,[2012](https://arxiv.org/html/2607.23648#bib.bib19)\), we divide the seeker’s emotional trajectory into three stages: the initial impact zone, the turbulence zone, and the integration zone\. This stage\-based design, compared with the fixed node approaches, allows natural emotion fluctuations within broader bounds\. In this way, emotional transitions no longer take the form of discrete label switching but instead constitute an evolution process with inherent logic\. See Appendix[A\.2](https://arxiv.org/html/2607.23648#A1.SS2)for stage definitions\.

Emotional trajectory control\.Guided by the stage model, this module determines the seeker’s current stage by analyzing the psychological structure, linguistic features, and emotional characteristics of the seeker’s utterances\. In real\-world psychological counseling scenarios, the seeker’s emotional development depends not only on their own willingness to change but also, to a large extent, on the counselor’s response style\(Ehrlichet al\.,[1979](https://arxiv.org/html/2607.23648#bib.bib20)\)\. Accordingly, this module further analyzes how the counselor’s response may influence the seeker’s subsequent reply and infers the next\-turn dynamics of emotional change\. Finally, based on the above analysis, the module outputs guidance for seeker’s emotional expression in the next turn, guiding the direction and intensity of emotional development within flexible stage boundaries, without enforcing deterministic emotional expression outcomes\.

#### 3\.2\.3Seeker Module

To avoid homogenized expression and unnatural dialogue caused by excessive activation of emotional schemas, we designed a schema activation mechanism based on conditional triggering and frequency constraints\.

Specifically, we maintain a long\-term and short\-term activation sequence for each type of emotional schema that every seeker possesses\. When a particular schema has been activated twice consecutively or five times in total, its activation is prohibited for the current turn\. Only when the schema is not subject to such restriction does the mechanism determine, based on specified conditions, whether the schema is permitted to be explicitly expressed in the current turn\. The pseudocode for this mechanism is provided in the[Algorithm 1](https://arxiv.org/html/2607.23648#alg1)\.

After completing the above determination, the module generates responses that exhibit stable personality and dynamic emotions based on the persona profile by integrating the emotional expression guidance provided by the ETC module for the current turn, as well as the schema activation status\.

Table 1:Statistics of Different Datasets\. NoT\., LoS\., LoC\. respectively represent average number of turns, average length of seeker’s response and average length of counselor’s response\.DatasetStatisticsNoT\.LoS\.LoC\.SMILECHAT10\.426\.128\.9CpsyCounD8\.730\.449\.7DeepWell\-Adol10\.028\.458\.5PsyDTCorpus18\.131\.658\.1EmoTrace\-D12\.234\.660\.3Table 2:Evaluation results for five datasets across three evaluation frameworks\. The best results are highlighted inbold, and the runner\-up results areunderlined\.Evaluation MatrixIndicatorDatasetsCpsyCounDSMILECHATDeepWell\-AdolPsyDTCorpusEmoTrace\-DCpsyCounComprehensiveness \(0–2\)1\.911\.8551\.9081\.951\.9Professionalism \(0–4\)2\.902\.443\.353\.943\.67Authenticity \(0–3\)2\.212\.302\.682\.932\.94Safety \(0–1\)0\.980\.98111PsyDTEmotional Empathy \(0–3\)1\.792\.012\.432\.82\.85Cognitive Empathy \(0–3\)2\.242\.392\.642\.772\.78Conversation Strategy \(0–3\)1\.971\.982\.472\.842\.80State and Attitude \(0–3\)2\.322\.312\.72\.812\.89Safety \(0–1\)0\.930\.940\.9911EmoTrace\-EEmotional Changes \(0–5\)2\.203\.002\.904\.214\.74Emotional Intensity andComplexity \(0–4\)1\.662\.602\.153\.033\.56Quality of Empathy \(0–5\)1\.832\.502\.754\.114\.49Autonomy \(0–4\)1\.712\.482\.313\.213\.73Motivation for Growth \(0–2\)1\.281\.571\.501\.931\.99
#### 3\.2\.4Counselor Module

A pipeline is designed in counselor module to generate responses in the following sequence: emotion analysis, intervention planning, and response generation\.

Analysis and planning submodule\.Grounded in EFT, this submodule divides the therapeutic process into three stages, namely emotion awareness, emotion deepening, and emotion transformation\. Based on the seeker’s utterances, the submodule analyzes the emotion type, intensity, triggers of the emotion and potential underlying emotional schemas\. On the basis of this analysis, the submodule selects an appropriate EFT therapeutic stage, formulates the current intervention goals and schema\-loosening strategies\. Finally, it integrates the above information to produce a detailed intervention strategy, providing a basis for subsequent response generation\.

Response generation submodule\.After receiving the analysis and planning results, this submodule is responsible for converting the structured intervention strategy into natural and coherent supportive language\. During generation, the model strictly adheres to the prescribed intervention plan and performs no additional emotion analysis or strategy adjustment, thereby ensuring consistency between the response and the planned intervention\.

### 3\.3Dialogue Generation

We use a role\-playing approach to generate multi\-turn psychological support dialogues by GPT\-4\.1\-mini based on persona profiles\. As shown in[Figure 2](https://arxiv.org/html/2607.23648#S1.F2), each turn involves interaction between seeker and counselor, while the ETC module generates guidance for the seeker’s emotional expression in the next turn based on the output of the current turn\. A complete dialogue is produced by iterating the above process\.

To ensure the completeness of emotional progression, we filter data by two criteria: full coverage of the seeker’s three emotional stages and reasonable stage duration to prevent excessive persistence\. Through this process, we construct EmoTrace\-D with 1,114 multi\-turn dialogues\. The related prompts and a complete dialogue case are presented in the Appendix[A\.2](https://arxiv.org/html/2607.23648#A1.SS2)\.

## 4General Dataset Quality Evaluation

To validate the quality of EmoTrace\-D, we compare it with four publicly available psychological support dialogue datasets: CpsyCounD\(Zhanget al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib6)\), SMILECHAT\(Qiuet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib5)\), DeepWell\-Adol\(Qiuet al\.,[2025](https://arxiv.org/html/2607.23648#bib.bib8)\), and PsyDTCorpus\(Xieet al\.,[2025](https://arxiv.org/html/2607.23648#bib.bib9)\)\.[Table 1](https://arxiv.org/html/2607.23648#S3.T1)presents the statistical information of these five datasets\.

![Refer to caption](https://arxiv.org/html/2607.23648v1/x3.png)Figure 3:Valence trajectories for five datasets\.To focus our evaluation on emotional richness and empathic specificity, we establish EmoTrace\-E to assess five dimensions: Emotional Variation, Emotional intensity and Complexity, Empathy Quality, Autonomy, and Growth Motivation\. To be fair, another two commonly used evaluation frameworks were also employed: the CpsyCoun evaluation matrix \(Comprehensiveness, Professionalism, Authenticity, Safety\), and the PsyDT evaluation matrix \(Affective Empathy, Cognitive Empathy, Dialogue strategy, State and Attitude, Safety\)\. All evaluations used a dual\-model evaluation strategy, scoring dataset quality with Deepseek\-V3\.2 and Gemini\-3\-flash respectively\. The average of the two scores was taken as the final result to mitigate bias introduced by any single evaluation model\. For detailed evaluation prompts, see[subsection B\.1](https://arxiv.org/html/2607.23648#A2.SS1)\.

![Refer to caption](https://arxiv.org/html/2607.23648v1/x4.png)Figure 4:Arousal trajectories for five datasets\.We randomly sample 100 samples from each of the five datasets\. As shown in[Table 2](https://arxiv.org/html/2607.23648#S3.T2)\. EmoTrace\-D outperforms the other four on most metrics, especially in emotional richness and empathic specificity\. Although its scores lower than PsyDTCorpus in Comprehensiveness and Professionalism, in real\-world psychological counseling scenarios, overly comprehensive event narration may reduce the proportion of emotional exploration within the dialogue\(Aleixoet al\.,[2021](https://arxiv.org/html/2607.23648#bib.bib21)\)\.

Furthermore, when counselors place excessive emphasis on professional psychological intervention techniques, it may compromise the seeker’s experience\. Therefore, a balance between professionalism and affinity needs to be considered\. As shown in the table, although EmoTrace\-D does not achieve the highest scores on these two dimensions, it still shows comparable or runner\-up performance, indicating that our dataset achieves balanced development across different dimensions\.

To show the effectiveness of EmoTrace\-D in enhancing emotional richness, we used the Valence\-Arousal \(VA\) model\(Russell,[1980](https://arxiv.org/html/2607.23648#bib.bib36)\)to plot the emotional trajectories of the five datasets\.[Figure 3](https://arxiv.org/html/2607.23648#S4.F3)is the valence trajectories, where the horizontal axis indicates dialogue progress and the vertical axis represents the positive negative degree of emotions\. Higher values reflect more positive emotions, and lower values more negative ones\. The trajectory of EmoTrace\-D shows the largest peak and the smallest valley compared to the other datasets, meaning it has the broadest emotional coverage and thus the highest emotional richness\.

[Figure 4](https://arxiv.org/html/2607.23648#S4.F4)presents the arousal trajectories where the horizontal axis is dialogue progress and the vertical axis is physiological activation intensity\. Larger values indicate stronger emotional responses\. The EmoTrace\-D trajectory has the largest fluctuation amplitude without being overly erratic, meaning it enhances emotional expression intensity while keeping fluctuations within a reasonable range and avoiding uncontrolled emotional intensity\. For more implementation details, see Appendix[B\.2](https://arxiv.org/html/2607.23648#A2.SS2)\.

![Refer to caption](https://arxiv.org/html/2607.23648v1/x5.png)Figure 5:Results of human evaluation for EmoTrace\-M and two baseline models\.
## 5Dataset\-Model Adaptation Experiment

### 5\.1Model Training

To investigate whether the dataset can effectively enhance the psychological counseling capabilities of LLMs, we fine\-tune Qwen3\-8B on EmoTrace\-D for 4 epochs, resulting in the EmoTrace\-M psychological support dialogue model\. The whole implementation is based on the LLaMAFactory\(Zhenget al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib22)\), with a batch size of 2 per GPU\. We employ a cosine\-type learning rate scheduler with warm\-up ratio set to 0\.05, an initial learning rate of 1\.0e\-4, and leverage 16\-bit half\-precision floating\-point to accelerate training\. The random seed is set to 42\. In the inference phase, all the LLMs adopt the following configuration:temperature= 0\.7,top\_p= 0\.9\. All implementations were conducted on a single NVIDIA A100 GPU\.

The fine\-tuned EmoTrace\-M is compared against the following baselines:

- •Open\-source: Qwen3\-8B\(Yanget al\.,[2025a](https://arxiv.org/html/2607.23648#bib.bib23)\); LLama3\.1\-8B\(Touvronet al\.,[2023](https://arxiv.org/html/2607.23648#bib.bib4)\); GLM4\-9B\(Glmet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib24)\)\.
- •Domain\-specific: CpsyCounX\(Zhanget al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib6)\); MeChat\(Qiuet al\.,[2024](https://arxiv.org/html/2607.23648#bib.bib5)\); MindChat\-Qwen\-7B\-v2\(Xin Yan,[2023](https://arxiv.org/html/2607.23648#bib.bib7)\)\.

Table 3:Model evaluation results in the automatic evaluation phase\. The best results are highlighted inbold, and the runner\-up results areunderlined\. The definitions of each indicator are as follows: Emotional Empathy \(Emo\.\), Conversation Strategy \(Con\.\), State and Attitude \(Sta\.\), Safe \(Saf\.\)\.TypeModelIndicatorEmo\. \(0–5\)Con\. \(0–5\)Sta\. \(0–3\)Saf\. \(0–2\)DomainCpsyCounX2\.311\.301\.361\.80MeChat2\.821\.831\.821\.71MindChat\-Qwen\-7B\-v22\.701\.461\.341\.76OpenLlama3\.1\-8B3\.352\.982\.201\.77GLM4\-9B3\.592\.852\.481\.91Qwen3\-8B3\.483\.742\.261\.78OursEmoTrace\-M4\.514\.182\.761\.94Table 4:Ablation study of different components\.ModelEmo\. \(0–5\)Emo\.& Com\. \(0–4\)Qua\. \(0–5\)Auto\. \(0–4\)Mot\. \(0–2\)Cog\. \(0–3\)w/o ETC4\.173\.124\.253\.241\.762\.81w/o Schema4\.503\.514\.713\.531\.912\.80all4\.783\.594\.613\.752\.002\.83
### 5\.2Evaluation Methods

We compare EmoTrace\-M with several open\-source general\-purpose models and domain\-specific psychological models to evaluate their dialogue capabilities in complex emotional scenarios\.

We adopt a simulated dialogue approach to construct evaluation samples for each model\. Specifically, GPT\-4\.1\-mini simulates the seeker based on 50 persona profiles that were not used in corpus generation\. To introduce a certain degree of emotional instability, this process retains emotional schemas and the schema activation mechanism\. The seeker engages in multi\-turn dialogues with each model, limited to 10 turns per dialogue\.

In the automatic evaluation phase, we also employ a dual\-model evaluation strategy\. To assess the models’ capabilities in emotion understanding and dialogue progression when simulating the counselor role, we reuse the PsyDT evaluation matrix with adaptations \(see the Appendix[B\.1](https://arxiv.org/html/2607.23648#A2.SS1)\)\.

For human evaluation, we selected the two models that achieved the highest scores among open\-source models and domain specific models respectively in the automatic evaluation phase, and compared them against EmoTrace\-M\. Four psychology experts and six psychology graduate students are invited to participate in the scoring process\. The evaluation criteria are identical to those used in the automatic evaluation phase\.

### 5\.3Results

[Table 3](https://arxiv.org/html/2607.23648#S5.T3)presents the score comparisons between our model and all baseline models\. The results demonstrate that our model achieves the highest scores across all metrics\. In terms of empathy quality and dialogue strategy, our model outperforms the general\-purpose baseline models\. We speculate the reason is, although general\-purpose models exhibit competitive dialogue capabilities, they tend to fall into patterns of shallow empathy or generic advice\-giving\. In contrast, our model effectively balances emotional support and strategic intervention\.

In addition, compared with previous domain\-specific models, our model still achieves a qualitative leap\. This indicates that models trained on datasets that focus solely on optimizing counselor responses or rely on static annotations struggle to handle complex emotional expressions from the seeker\. In contrast, by extending the focus of optimization from the counselor to the seeker, our approach enables the model to learn strategies for handling emotional instability and underlying cognitive structures, resulting in more professional and empathetic counseling dialogues\. A detailed case study is provided in the Appendix[C](https://arxiv.org/html/2607.23648#A3)\.

[Figure 5](https://arxiv.org/html/2607.23648#S4.F5)presents the results of the manual evaluation, showing that EmoTrace\-M achieved the highest scores across all four dimensions, consistent with the results of the automated evaluation\.

### 5\.4Ablation Experiment

To further investigate the respective contributions of the ETC module and the emotional schema activation mechanism in our framework, we conduct an ablation study comparing the full model \(all\) with two variants: \(1\) w/o ETC and \(2\) w/o schema\. To deeply examine the impact of the core designs on cognitive structures, we add one additional metric to the original evaluation framework \(see[subsection B\.1](https://arxiv.org/html/2607.23648#A2.SS1)\)\. The results are presented in[Table 4](https://arxiv.org/html/2607.23648#S5.T4)\.

After removing ETC, a significant decline is observed in both Emotional Variation and Emotional Intensity and Complexity\. Disabling the emotional schema activation mechanism also leads to declines across multiple metrics, albeit to a lesser extent compared to removing ETC, suggesting that under the guidance of ETC, the seeker remains capable of simulating complex emotions with a certain degree of cognitive patterning\.

An interesting observation is that when the schema activation mechanism is removed, the Empathy Quality score increases slightly\. We hypothesize that this occurs because, in the absence of complex cognitive defense mechanisms on the seeker’s part, the counselor model can more easily achieve higher scores through shallow empathic responses\. In contrast, although our full model \(all\) yields a slightly lower Empathy Quality score, it more faithfully reflects the real\-world challenge in which counselors must contend with the seeker’s internal cognitive resistance\.

## 6Conclusion

This paper proposes EmoTrace, a multi\-turn psychological support dialogue generation framework centered on the seeker’s emotional trajectory\. It integrates the seeker module with persona profile and emotional schema activation mechanism, an EFT driven counselor module, and an emotional trajectory control module to generate more authentic, complex, and layered dialogues\. Experiments show that the dataset and dialogue model constructed using this method outperform existing approaches in emotional richness, empathy quality, and the capacity to respond to complex emotional states\.

## Limitations

Although the experimental results demonstrate the effectiveness of EmoTrace, several limitations warrant further attention\. In real\-world settings, psychological counseling is a complex endeavor, and the counseling process may integrate multiple therapeutic modalities\. Our framework is primarily designed around Emotion\-Focused Therapy\. Thus, how to extend and accommodate other schools of psychological counseling remains an open question\. Furthermore, the dialogue synthesis process involves information transfer across multiple modules, which significantly increases generation costs and consequently limits the scalability of the dataset\. Future research should consider how to reduce generation costs as much as possible while preserving output quality\.

## Ethical Statement

During the persona profile construction and dialogue generation processes, we implemented a rigorous data cleaning pipeline, including rule\-based filtering, manual rewriting, and manual proofreading, to ensure that the final dataset contained no personally identifiable information or sensitive content\. Additionally, we removed any dialogue content that could potentially cause harm to the seeker, others, or society, thereby mitigating potential risks\.

During the evaluation phase, we conducted strict safety assessments on both the dataset and the model outputs\. However, the model’s response generation process lacks human intervention and feedback\. Moreover, given the substantial variability in the psychological conditions of different users, certain responses may inevitably cause harm to some users\. Therefore, EmoTrace\-M is intended solely as a supplementary counseling tool and cannot replace genuine psychological therapy\. Users experiencing severe psychological distress should seek timely assistance from professional counselors or psychiatrists\. Furthermore, when deploying this model in downstream applications, it is mandatory to inform users in advance that the responses generated by the AI model should be used only as a reference\.

## References

- J\. Achiam, S\. Adler, S\. Agarwal, L\. Ahmad, I\. Akkaya, F\. L\. Aleman, D\. Almeida, J\. Altenschmidt, S\. Altman, S\. Anadkat,et al\.\(2023\)Gpt\-4 technical report\.arXiv preprint arXiv:2303\.08774\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1)\.
- A review of empirical studies investigating narrative, emotion and meaning\-making modes and client process markers in psychotherapy\.Journal of Contemporary Psychotherapy51\(1\),pp\. 31–40\.Cited by:[§4](https://arxiv.org/html/2607.23648#S4.p3.1)\.
- A\. T\. Beck, A\. J\. Rush, B\. F\. Shaw, G\. Emery, R\. J\. DeRubeis, and S\. D\. Hollon \(2024\)Cognitive therapy of depression\.Guilford Publications\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p4.1),[§3\.2\.1](https://arxiv.org/html/2607.23648#S3.SS2.SSS1.p2.1)\.
- Y\. Chen, X\. Xing, J\. Lin, H\. Zheng, Z\. Wang, Q\. Liu, and X\. Xu \(2023\)Soulchat: improving llms’ empathy, listening, and comfort abilities through fine\-tuning with multi\-turn empathy conversations\.InFindings of the Association for Computational Linguistics: EMNLP 2023,pp\. 1170–1183\.Cited by:[§2\.2](https://arxiv.org/html/2607.23648#S2.SS2.p1.1)\.
- R\. P\. Ehrlich, A\. R\. D’Augelli, and S\. J\. Danish \(1979\)Comparative effectiveness of six counselor verbal responses\.\.Journal of Counseling Psychology26\(5\),pp\. 390\.Cited by:[§3\.2\.2](https://arxiv.org/html/2607.23648#S3.SS2.SSS2.p4.1)\.
- T\. Glm, A\. Zeng, B\. Xu, B\. Wang, C\. Zhang, D\. Yin, D\. Zhang, D\. Rojas, G\. Feng, H\. Zhao,et al\.\(2024\)Chatglm: a family of large language models from glm\-130b to glm\-4 all tools\.arXiv preprint arXiv:2406\.12793\.Cited by:[1st item](https://arxiv.org/html/2607.23648#S5.I1.i1.p1.1)\.
- L\. S\. Greenberg \(2004\)Emotion–focused therapy\.Clinical Psychology & Psychotherapy: An International Journal of Theory & Practice11\(1\),pp\. 3–16\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p5.1)\.
- L\. S\. Greenberg \(2010\)Emotion\-focused therapy: a clinical synthesis\.Focus8\(1\),pp\. 32–42\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p5.1)\.
- L\. S\. Greenberg \(2012\)Emotions, the great captains of our lives: their role in the process of change in psychotherapy\.\.American Psychologist67\(8\),pp\. 697\.Cited by:[§3\.2\.2](https://arxiv.org/html/2607.23648#S3.SS2.SSS2.p3.1)\.
- H\. Hu, C\. Ma, Q\. Wang, L\. Lin, Y\. Zhou, L\. Cui, F\. Ma, and Q\. Tian \(2026\)Theramind: a strategic and adaptive agent for longitudinal psychological counseling\.InProceedings of the ACM Web Conference 2026,pp\. 9136–9147\.Cited by:[§2\.2](https://arxiv.org/html/2607.23648#S2.SS2.p2.1)\.
- R\. L\. Leahy \(2002\)A model of emotional schemas\.Cognitive and behavioral practice9\(3\),pp\. 177–190\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p4.1),[§3\.2\.1](https://arxiv.org/html/2607.23648#S3.SS2.SSS1.p2.1)\.
- S\. Lee, S\. Mac Kim, M\. Kim, D\. Kang, D\. Yang, H\. Kim, M\. Kang, D\. Jung, M\. H\. Kim, S\. Lee,et al\.\(2024\)Cactus: towards psychological counseling conversations using cognitive behavioral theory\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 14245–14274\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p2.1)\.
- S\. Liu, C\. Zheng, O\. Demasi, S\. Sabour, Y\. Li, Z\. Yu, Y\. Jiang, and M\. Huang \(2021\)Towards emotional support dialog systems\.InProceedings of the 59th annual meeting of the association for computational linguistics and the 11th international joint conference on natural language processing \(volume 1: Long papers\),pp\. 3469–3483\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p4.1)\.
- T\. Mayer, N\. Warikoo, O\. Grimm, A\. Reif, and I\. Gurevych \(2022\)GDPR compliant collection of therapist\-patient\-dialogues\.arXiv preprint arXiv:2211\.12360\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1)\.
- M\. Prince, V\. Patel, S\. Saxena, M\. Maj, J\. Maselko, M\. R\. Phillips, and A\. Rahman \(2007\)No health without mental health\.The lancet370\(9590\),pp\. 859–877\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p1.1)\.
- Z\. Qi, T\. Kaneko, K\. Takamizo, M\. Ukiyo, and M\. Inaba \(2025\)Kokorochat: a japanese psychological counseling dialogue dataset collected via role\-playing by trained counselors\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 12424–12443\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1)\.
- H\. Qiu, H\. He, S\. Zhang, A\. Li, and Z\. Lan \(2024\)Smile: single\-turn to multi\-turn inclusive language expansion via chatgpt for mental health support\.InFindings of the Association for Computational Linguistics: EMNLP 2024,pp\. 615–636\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p1.1),[§4](https://arxiv.org/html/2607.23648#S4.p1.1),[2nd item](https://arxiv.org/html/2607.23648#S5.I1.i2.p1.1)\.
- W\. Qiu, Y\. Wang, J\. Tan, H\. Hou, Q\. Liu, W\. Yao, and S\. Ni \(2025\)DeepWell\-adol: a scalable expert\-based dialogue corpus for adolescent positive mental health and wellbeing promotion\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 12797–12821\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p3.1),[§4](https://arxiv.org/html/2607.23648#S4.p1.1)\.
- J\. A\. Russell \(1980\)A circumplex model of affect\.\.Journal of personality and social psychology39\(6\),pp\. 1161\.Cited by:[§4](https://arxiv.org/html/2607.23648#S4.p5.1)\.
- Y\. Shi, L\. Zhang, M\. Li, Y\. Zheng, X\. Wang, and F\. Kong \(2026\)Simulating human\-like counseling: a path\-and scenario\-guided framework for psychological support dialogue\.InProceedings of the AAAI Conference on Artificial Intelligence,Vol\.40,pp\. 17688–17696\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p2.1)\.
- V\. Sorin, D\. Brin, Y\. Barash, E\. Konen, A\. Charney, G\. Nadkarni, and E\. Klang \(2024\)Large language models and empathy: systematic review\.Journal of medical Internet research26,pp\. e52597\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1)\.
- W\. B\. Stiles, R\. Elliott, S\. P\. Llewelyn, J\. A\. Firth\-Cozens, F\. R\. Margison, D\. A\. Shapiro, and G\. Hardy \(1990\)Assimilation of problematic experiences by clients in psychotherapy\.\.Psychotherapy: Theory, Research, Practice, Training27\(3\),pp\. 411\.Cited by:[§3\.2\.2](https://arxiv.org/html/2607.23648#S3.SS2.SSS2.p2.1)\.
- H\. Sun, Z\. Lin, C\. Zheng, S\. Liu, and M\. Huang \(2021\)Psyqa: a chinese dataset for generating long counseling text for mental health support\.InFindings of the association for computational linguistics: ACL\-IJCNLP 2021,pp\. 1489–1503\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p1.1)\.
- H\. Touvron, T\. Lavril, G\. Izacard, X\. Martinet, M\. Lachaux, T\. Lacroix, B\. Rozière, N\. Goyal, E\. Hambro, F\. Azhar,et al\.\(2023\)Llama: open and efficient foundation language models\.arXiv preprint arXiv:2302\.13971\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1),[1st item](https://arxiv.org/html/2607.23648#S5.I1.i1.p1.1)\.
- T\. J\. VanderWeele, B\. R\. Johnson, M\. Bradshaw, D\. M\. Goodman, L\. D\. Kubzansky, T\. Lomas, A\. Moreira\-Almeida, C\. N\. Okafor, S\. T\. Ouyang, and V\. Patel \(2026\)Mental illness, mental health, and mental well\-being\.Npj mental health research5\(1\),pp\. 11\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p1.1)\.
- N\. Wang, Z\. Peng, H\. Que, J\. Liu, W\. Zhou, Y\. Wu, H\. Guo, R\. Gan, Z\. Ni, J\. Yang,et al\.\(2024\)Rolellm: benchmarking, eliciting, and enhancing role\-playing abilities of large language models\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 14743–14777\.Cited by:[§3\.2\.1](https://arxiv.org/html/2607.23648#S3.SS2.SSS1.p1.1)\.
- X\. Wang, W\. Shi, R\. Kim, Y\. Oh, S\. Yang, J\. Zhang, and Z\. Yu \(2019\)Persuasion for good: towards a personalized persuasive dialogue system for social good\.InProceedings of the 57th annual meeting of the association for computational linguistics,pp\. 5635–5649\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p1.1)\.
- A\. Welivita and P\. Pu \(2022\)Curating a large\-scale motivational interviewing dataset using peer support forums\.InProceedings of the 29th International Conference on Computational Linguistics,pp\. 3315–3330\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p1.1)\.
- Y\. Wu, G\. Wan, J\. Li, S\. Zhao, L\. Ma, T\. Ye, M\. Zhang, I\. Pop, Y\. Zhang, and J\. Chen \(2026\)WiseMind: a knowledge\-guided multi\-agent framework for accurate and empathetic psychiatric diagnosis\.npj Digital Medicine\.Cited by:[§2\.2](https://arxiv.org/html/2607.23648#S2.SS2.p2.1)\.
- H\. Xie, Y\. Chen, X\. Xing, J\. Lin, and X\. Xu \(2025\)Psydt: using llms to construct the digital twin of psychological counselor with personalized counseling style for psychological counseling\.InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\),pp\. 1081–1115\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p3.1),[§4](https://arxiv.org/html/2607.23648#S4.p1.1)\.
- D\. X\. Xin Yan \(2023\)MindChat: psychological large language model\.GitHub\.Note:[https://github\.com/X\-D\-Lab/MindChat](https://github.com/X-D-Lab/MindChat)Cited by:[§2\.2](https://arxiv.org/html/2607.23648#S2.SS2.p1.1),[2nd item](https://arxiv.org/html/2607.23648#S5.I1.i2.p1.1)\.
- A\. Yang, A\. Li, B\. Yang, B\. Zhang, B\. Hui, B\. Zheng, B\. Yu, C\. Gao, C\. Huang, C\. Lv,et al\.\(2025a\)Qwen3 technical report\.arXiv preprint arXiv:2505\.09388\.Cited by:[1st item](https://arxiv.org/html/2607.23648#S5.I1.i1.p1.1)\.
- T\. Yang, Y\. Zhu, X\. Quan, C\. Liu, and Q\. Wang \(2025b\)Psyplay: personality\-infused role\-playing conversational agents\.arXiv preprint arXiv:2502\.03821\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p4.1)\.
- C\. Zhang, R\. Li, M\. Tan, M\. Yang, J\. Zhu, D\. Yang, J\. Zhao, G\. Ye, C\. Li, and X\. Hu \(2024\)Cpsycoun: a report\-based multi\-turn dialogue reconstruction and evaluation framework for chinese psychological counseling\.InFindings of the Association for Computational Linguistics: ACL 2024,pp\. 13947–13966\.Cited by:[§1](https://arxiv.org/html/2607.23648#S1.p2.1),[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p1.1),[§4](https://arxiv.org/html/2607.23648#S4.p1.1),[2nd item](https://arxiv.org/html/2607.23648#S5.I1.i2.p1.1)\.
- Y\. Zheng, R\. Zhang, J\. Zhang, Y\. Ye, and Z\. Luo \(2024\)Llamafactory: unified efficient fine\-tuning of 100\+ language models\.InProceedings of the 62nd annual meeting of the association for computational linguistics \(volume 3: system demonstrations\),pp\. 400–410\.Cited by:[§5\.1](https://arxiv.org/html/2607.23648#S5.SS1.p1.1)\.
- J\. Zhou, Y\. Chen, J\. Yin, Y\. Huang, Y\. Shi, X\. Zhang, L\. Peng, R\. Zhang, T\. Lv, Z\. Hu,et al\.\(2025a\)Crisp: cognitive restructuring of negative thoughts through multi\-turn supportive dialogues\.InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing,pp\. 32462–32491\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p2.1)\.
- Y\. Zhou, N\. Zhou, Q\. Chen, J\. Zhou, A\. Zhou, and L\. He \(2025b\)DiaCBT: a long\-periodic dialogue corpus guided by cognitive conceptualization diagram for cbt\-based psychological counseling\.arXiv preprint arXiv:2509\.02999\.Cited by:[§2\.1](https://arxiv.org/html/2607.23648#S2.SS1.p2.1)\.

## Appendix ADetails of EmoTrace Framework

![Refer to caption](https://arxiv.org/html/2607.23648v1/x6.png)Figure 6:Topic distribution of EmoTrace\-D\.### A\.1Persona Profile Construction

Table 5:Emotional Schema Definitions and Examples\.AxisSchemaDefinitionExampleSelfRational SupremacyA tendency to suppress or analyze emotions through rationality, believing that emotions should not dominate behavior\.“I know I should allow myself to feel disappointed, but I feel that this kind of negative emotion simply shouldn’t exist\.”GuiltA tendency to feel self\-reproach when one’s own actions or outcomes negatively affect others\.“It’s all because of what I said that day that he’s suffering so much now—I should never have opened my mouth\.”ShameA tendency to regard one’s own emotions or circumstances as deficiencies, accompanied by concerns about being evaluated or rejected by others\.“Actually, I’ve been in a bad state lately… but voicing it would only make me seem fragile and weak, wouldn’t it?”OthersOthers’ IncomprehensionA belief that others find it difficult to understand or accept one’s emotions and circumstances\.“There’s no point in explaining it to them\. They’ll never understand what it feels like to have everyone staring at you\.”Loss of ControlA concern that once expressed, emotions may become uncontrollable and potentially disrupt relationships or situations\.“If I let my anger out right now, the situation would completely spiral, and no one would be able to get it under control\.”ComplianceA tendency to suppress one’s own emotions to meet others’ expectations or maintain relational stability\.“It’s fine, just decide among yourselves\. I’m okay with anything\. I don’t want things to become unpleasant because of me\.”World/FutureProlonged DurationA belief that negative emotions will persist for a long time and are difficult to end or alleviate\.“This kind of low mood doesn’t just go away in a few days like it does for other people\. Once it sets in, it clings to me for a very, very long time\.”Philosophical ReflectionA tendency to understand and interpret emotions in abstract terms, through the lens of life meaning or existential dimensions\.“In the end, this kind of anxiety is merely the price one has to bear when facing nothingness\.”Table 6:Definitions of Three\-stage Emotional Trajectory\.StagePsychological Structure CharacteristicsLinguistic Expression CharacteristicsCore Emotional SpectrumInitial Impact•The seeker has just encountered the problem or stressful event triggering the emotion\.•Prominent psychological defenses are present\.•Cognitive and emotional content exists but has not yet been clearly perceived or expressed\.•A tendency to attribute problems to external factors or other people\.•Events are described more than emotions\.•Emotional expression is vague or indirect\.•Few emotion words are used\.Emotions at this stage are largely inhibitory secondary emotions\. The primary types include confusion, anxiety, unease, grievance, and irritability\.Turbulence•Emotional experience is markedly intensified\.•Internal conflicts begin to re\-emerge\.•Emotions may fluctuate or recur\.•Emotional schemas begin to be activated frequently\.•Emotional expression becomes more direct, intense, or contradictory\.•Narratives contain abundant emotion words\.•Self\-evaluation or self\-blame emerges\.Emotions are more primitive and intense, though not yet integrated\. The primary types include sadness, anger, fear, shame, loneliness, and disappointment\.Integration•A clearer understanding of emotions is achieved\.•The ability to connect emotions, needs, and behaviors emerges\.•Internal conflicts diminish, and self\-acceptance increases\.•Expression becomes calmer and more coherent\.•Reflective language appears\.•Greater focus is placed on the future or on change\.Transformative emotions emerge\. The primary types include relief, acceptance, hope, and calmness\.![Refer to caption](https://arxiv.org/html/2607.23648v1/x7.png)Figure 7:The prompt used for seeker persona profile generation\.[Table 5](https://arxiv.org/html/2607.23648#A1.T5)presents the complete definitions of the emotional schemas\.[Figure 7](https://arxiv.org/html/2607.23648#A1.F7)shows the prompt used for seeker persona profile generation\.

### A\.2Multi\-Turn Psychological Support Dialogues Synthesis

![Refer to caption](https://arxiv.org/html/2607.23648v1/x8.png)Figure 8:The prompt used for the seeker module\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x9.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x10.png)Figure 9:The prompt used for the analysis and planning submodule in the counselor module\.![Refer to caption](https://arxiv.org/html/2607.23648v1/x11.png)Figure 10:The prompt used for the generation submodule in the counselor module\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x12.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x13.png)Figure 11:The prompt used for the ETC module\.![Refer to caption](https://arxiv.org/html/2607.23648v1/x14.png)Figure 12:A complete example of a multi\-turn dialogue corpus\.The definitions of the three\-stage emotional trajectory are provided in[Table 6](https://arxiv.org/html/2607.23648#A1.T6)\. The prompts for the seeker module, the counselor module \(including the analysis and planning submodule and the generation submodule\), and the ETC module are shown in[Figure 8](https://arxiv.org/html/2607.23648#A1.F8),[9](https://arxiv.org/html/2607.23648#A1.F9),[10](https://arxiv.org/html/2607.23648#A1.F10), and[11](https://arxiv.org/html/2607.23648#A1.F11), respectively\.[Figure 12](https://arxiv.org/html/2607.23648#A1.F12)presents a complete example of a multi\-turn dialogue corpus\.

Finally, we constructed EmoTrace\-D, which covers 12 topics, including emotion, family, interpersonal relationship, therapy, marriage, growth, self, behavior, society, workplace, sexual psychology, and psychological knowledge\. The distribution of these topics is shown in[Figure 6](https://arxiv.org/html/2607.23648#A1.F6)\.

## Appendix BDetails of Experiments

### B\.1Evaluation prompt

![Refer to caption](https://arxiv.org/html/2607.23648v1/x15.png)Figure 13:The prompt of CpsyCoun evaluation matrix\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x16.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x17.png)Figure 14:The prompt of PsyDT evaluation matrix\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x18.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x19.png)Figure 15:The prompt of EmoTrace\-E\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x20.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x21.png)Figure 16:The prompt for the evaluation metrics used in model evaluation\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x22.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x23.png)Figure 17:The prompt for the evaluation metrics used in ablation experiment\.The prompts for the three evaluation frameworks used in dataset evaluation are presented in[Figure 13](https://arxiv.org/html/2607.23648#A2.F13),[14](https://arxiv.org/html/2607.23648#A2.F14), and[15](https://arxiv.org/html/2607.23648#A2.F15)\. The prompts for the evaluation metrics used in model evaluation and ablation experiment are shown in[Figure 16](https://arxiv.org/html/2607.23648#A2.F16)and[17](https://arxiv.org/html/2607.23648#A2.F17)\.

### B\.2Visualization of Emotional Trajectories

To construct visualizable emotional trajectories, we manually annotated Valence and Arousal for each turn of the seeker’s utterances\. Since different dialogues contain varying numbers of turns, direct averaging across turns is not feasible\. Therefore, we mapped each dialogue onto a unified normalized time axis\. For a dialogue consisting ofNNturns, the dialogue progress of theii\-th turn is defined as:

ti=i−1N−1,i=1,2,…,Nt\_\{i\}=\\frac\{i\-1\}\{N\-1\},\\quad i=1,2,\\ldots,N\(1\)
thereby uniformly mapping all dialogues to the interval\[0,1\]\[0,1\]\. Subsequently, we employed Piecewise Cubic Hermite Interpolating Polynomial \(PCHIP\) to interpolate the discrete emotional points, resampling each trajectory into a continuous trajectory of fixed length \(set to 20 uniformly distributed sampling points in this experiment\)\. After resampling, we computed the mean and standard deviation for both Valence and Arousal across all dialogues at each normalized progress point\.

## Appendix CCase Study

![Refer to caption](https://arxiv.org/html/2607.23648v1/x24.png)Figure 18:A case of a seeker confiding to MeChat\.![Refer to caption](https://arxiv.org/html/2607.23648v1/x25.png)Figure 19:A case of a seeker confiding to Qwen3\.![[Uncaptioned image]](https://arxiv.org/html/2607.23648v1/x26.png)![Refer to caption](https://arxiv.org/html/2607.23648v1/x27.png)Figure 20:A case of a seeker confiding to EmoTrace\-M\.In this subsection, we present case studies in which MeChat, Qwen3, and EmoTrace\-M are used to simulate counseling dialogues between a counselor and a seeker, as illustrated in[Figure 18](https://arxiv.org/html/2607.23648#A3.F18),[19](https://arxiv.org/html/2607.23648#A3.F19), and[20](https://arxiv.org/html/2607.23648#A3.F20)\.

When acting as the counselor, MeChat tends to provide reassurance and value affirmation without sufficiently exploring the seeker’s specific emotional experiences\. In addition, it offers intervention strategies before allowing the seeker to fully express their internal state\. Such jumps in conversational flow may cause the seeker to feel instructed rather than accompanied and understood\. The main limitation of Qwen3 lies in introducing psychological terminology and structured techniques too early and too frequently, which weakens empathic attunement to the seeker’s emotional state itself\.

By contrast, during the early stage of counseling, EmoTrace\-M proactively guides the seeker to attend to bodily reactions associated with emotions, helping transform diffuse anxiety into a concrete experience that can be explored\. Throughout the dialogue, EmoTrace\-M does not rush to explain or intervene\. Instead, it first normalizes and de\-shames the seeker’s feelings, while inviting collaborative exploration of the underlying causes of those feelings\. When the seeker is able to articulate core distress, EmoTrace\-M can accurately identify the underlying motivations and support meaning reconstruction\.

Algorithm 1Schema Activation Constraint Mechanism0:Persona profile

PPcontaining emotion schemas

0:Activation constraint

1:

schemas←P​\[“emotion schemas”\]\\textit\{schemas\}\\leftarrow P\[\\text\{\`\`emotion schemas''\}\]
2:

activation\_dict←\{s:\[\]​for each schema​s∈schemas\}\\textit\{activation\\\_dict\}\\leftarrow\\\{\\,s:\[\\,\]\\text\{ for each schema \}s\\in\\textit\{schemas\}\\,\\\}
3:

turn←1\\textit\{turn\}\\leftarrow 1
4:whiledialogue is not finisheddo

5:if

turn=1\\textit\{turn\}=1then

6:

activation\_constraint←∅\\textit\{activation\\\_constraint\}\\leftarrow\\emptyset
7:else

8:

forbidden\_schemas←∅\\textit\{forbidden\\\_schemas\}\\leftarrow\\emptyset
9:for all

\(s,l\)\(s,l\)inactivation\_dictdo

10:

c1←Sum​\(l\)≥5c\_\{1\}\\leftarrow\\textsc\{Sum\}\(l\)\\geq 5
11:

c2←\(\|l\|≥2∧l​\[−1\]=1∧l​\[−2\]=1\)c\_\{2\}\\leftarrow\(\|l\|\\geq 2\\land l\[\-1\]=1\\land l\[\-2\]=1\)
12:if

c1∨c2c\_\{1\}\\lor c\_\{2\}then

13:

forbidden\_schemas←forbidden\_schemas∪\{s\}\\textit\{forbidden\\\_schemas\}\\leftarrow\\textit\{forbidden\\\_schemas\}\\cup\\\{s\\\}
14:endif

15:endfor

16:

activation\_constraint←forbidden\_schemas\\textit\{activation\\\_constraint\}\\leftarrow\\textit\{forbidden\\\_schemas\}
17:endif

18:

seeker\_output←Seeker​\(activation\_constraint\)\\textit\{seeker\\\_output\}\\leftarrow\\textsc\{Seeker\}\(\\textit\{activation\\\_constraint\}\)
19:

activation←seeker\_output​\[“schema\_activation”\]\\textit\{activation\}\\leftarrow\\textit\{seeker\\\_output\}\[\\text\{\`\`schema\\\_activation''\}\]
20:

activated←activation​\[“activated”\]\\textit\{activated\}\\leftarrow\\textit\{activation\}\[\\text\{\`\`activated''\}\]
21:

schema\_name←activation​\[“schema\_name”\]\\textit\{schema\\\_name\}\\leftarrow\\textit\{activation\}\[\\text\{\`\`schema\\\_name''\}\]
22:for all

ssinschemasdo

23:if

activated∧schema\_name=s\\textit\{activated\}\\land\\textit\{schema\\\_name\}=sthen

24:

activation\_dict​\[s\]\.Append​\(1\)\\textit\{activation\\\_dict\}\[s\]\.\\textsc\{Append\}\(1\)
25:else

26:

activation\_dict​\[s\]\.Append​\(0\)\\textit\{activation\\\_dict\}\[s\]\.\\textsc\{Append\}\(0\)
27:endif

28:endfor

29:

turn←turn\+1\\textit\{turn\}\\leftarrow\\textit\{turn\}\+1
30:endwhile