Paper deep dive
STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems
Hongru Ji, Yuyin Fan, Meng Zhao, Xianghua Li, Lianwei Wu, Chao Gao
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 4/10/2026, 3:48:09 AM
Summary
STRIDE-ED is a strategy-grounded, interpretable reasoning framework for empathetic dialogue systems. It addresses limitations in existing models by introducing a comprehensive empathy strategy system, a multi-stage cognitive reasoning process (CoT), and a high-quality data refinement pipeline. The framework utilizes LLM-based annotation, consistency-weighted evaluation, and a two-stage training paradigm (supervised fine-tuning followed by multi-objective reinforcement learning) to improve emotional alignment and strategic response generation.
Entities (5)
Relation Signals (3)
STRIDE-ED ā employs ā DeepSeek R1
confidence 95% Ā· we employ DeepSeek-R1, a large language model with strong reasoning capabilities, as an automated annotation expert.
STRIDE-ED ā uses ā Proximal Policy Optimization
confidence 95% Ā· we apply Proximal Policy Optimization (PPO) to optimize the model towards outputs
STRIDE-ED ā utilizes ā EMPATHETICDIALOGUES
confidence 95% Ā· we automatically annotate the EMPATHETICDIALOGUES dataset (Rashkin et al., 2019) with strategy types and rationales
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Empathetic dialogue requires not only recognizing a user's emotional state but also making strategy-aware, context-sensitive decisions throughout response generation. However, the lack of a comprehensive empathy strategy framework, explicit task-aligned multi-stage reasoning, and high-quality strategy-aware data fundamentally limits existing approaches, preventing them from effectively modeling empathetic dialogue as a complex, multi-stage cognitive and decision-making process. To address these challenges, we propose STRIDE-ED, a STRategy-grounded, Interpretable, and DEep reasoning framework that models Empathetic Dialogue through structured, strategy-conditioned reasoning. To support effective learning, we develop a strategy-aware data refinement pipeline integrating LLM-based annotation, multi-model consistency-weighted evaluation, and dynamic sampling to construct high-quality training data aligned with empathetic strategies. Furthermore, we adopt a two-stage training paradigm that combines supervised fine-tuning with multi-objective reinforcement learning to better align model behaviors with target emotions, empathetic strategies, and response formats. Extensive experiments demonstrate that STRIDE-ED generalizes across diverse open-source LLMs and consistently outperforms existing methods on both automatic metrics and human evaluations.
Tags
Links
- Source: https://arxiv.org/abs/2604.07100v1
- Canonical: https://arxiv.org/abs/2604.07100v1
Trouble viewing inline? Open PDF directly ā
Full Text
58,526 characters extracted from source content.
Expand or collapse full text
STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems Hongru Ji 1 , Yuyin Fan 1 , Meng Zhao 2 , Xianghua Li 1 , Lianwei Wu 1 , Chao Gao 1 1 Northwestern Polytechnical University, 2 Henan University of Technology Abstract Empathetic dialogue requires not only recog- nizing a userās emotional state but also mak- ing strategy-aware, context-sensitive decisions throughout response generation.However, the lack of a comprehensive empathy strategy framework, explicit task-aligned multi-stage reasoning, and high-quality strategy-aware data fundamentally limits existing approaches, pre- venting them from effectively modeling em- pathetic dialogue as a complex, multi-stage cognitive and decision-making process. To ad- dress these challenges, we propose STRIDE- ED, a STRategy-grounded, Interpretable, and DEep reasoning framework that models Em- pathetic Dialogue through structured, strategy- conditioned reasoning. To support effective learning, we develop a strategy-aware data re- finement pipeline integrating LLM-based anno- tation, multi-model consistency-weighted eval- uation, and dynamic sampling to construct high- quality training data aligned with empathetic strategies. Furthermore, we adopt a two-stage training paradigm that combines supervised fine-tuning with multi-objective reinforcement learning to better align model behaviors with target emotions, empathetic strategies, and re- sponse formats. Extensive experiments demon- strate that STRIDE-ED generalizes across di- verse open-source LLMs and consistently out- performs existing methods on both automatic metrics and human evaluations. Our data and code are available athttps://github.com/ jicoder-nwpu/STRIDE-ED. 1 Introduction Empathetic dialogue, a cornerstone of human so- cial interaction, requires not only the recognition of anotherās emotional state but also the formu- lation of a response that conveys understanding, validation, and appropriate support (Batson, 2009). From a psychological perspective, this is a com- plex, multi-stage cognitive and decision-making process (Davis, 1983; Gao et al., 2023; Xu and I was going to go to the movies with a bunch of friends. But they all bailed on me and now I have to go alone. I was going to go to the movies with a bunch of friends. But they all bailed on me and now I have to go alone. If the movie is worth it you will never miss them.If the movie is worth it you will never miss them.If the movie is worth it you will never miss them. Strategy SelectionStrategy Selection Cognitive ReframingCognitive Reframing Strategy Selection Cognitive Reframing I would apply Cognitive Reframing by shifting focus to the potential value of the movie itself, highlighting how enjoying the film can create fulfillment even in solitude, thereby easing the emphasis on missing others. Execution ActionsExecution Actions I would apply Cognitive Reframing by shifting focus to the potential value of the movie itself, highlighting how enjoying the film can create fulfillment even in solitude, thereby easing the emphasis on missing others. Execution Actions Emotion RecognitionEmotion Recognition lonelylonely Emotion Recognition lonely Emotion Recognition lonely Speaker planned to attend a movie with friends but now must go alone after cancellations. Scenario Summary Figure 1: Illustration of the STRIDE-ED reasoning pro- cess. Given a user utterance, the framework performs scenario summarization, emotion recognition, strategy inference, and strategy-guided response generation. Jiang, 2024). This highlights the strategic and context-sensitive nature of empathy, making empa- thetic dialogue a fundamentally challenging task beyond surface-level text generation. Early studies focused on implicitly enhancing dialogue models through external commonsense knowledge or affective lexicons. For example, prior works (Liu et al., 2022; Cai et al., 2023) incor- porates emotional commonsense graphs or com- monsense knowledge selection into neural architec- tures to provide latent cues for more emotion-aware generation. However, without explicit modeling of reasoning or decision processes (Zhang et al., 2024), these approaches offer limited insight into the mechanisms through which emotional under- standing informs response generation. To address this lack of transparency, recent research has turned to Large Language Models (LLMs) (Dubey et al., 2024; Liu et al., 2024; Yang et al., 2025), which enable more explicit model- arXiv:2604.07100v1 [cs.CL] 8 Apr 2026 ing of intermediate reasoning processes through Chain-of-Thought (CoT) prompting. Building on this direction, prior work such as (Lee et al., 2023; Hu et al., 2025) further encourages LLMs to gen- erate explicit intermediate rationales before pro- ducing responses, thereby enhancing interpretabil- ity. However, a fundamental limitation persists in these CoT-based approaches: the reasoning process lacks grounding in a rigorous strategic framework. Existing strategies (Liu et al., 2021; Zhang et al., 2025) often derived from specific domains such as psychological support, primarily address negative emotions and are confined to low-level responses. They fail to encompass the full emotional spec- trum, particularly positive and neutral states, and lack provisions for higher-order cognitive strate- gies, as shown in Figure 1. Consequently, while these existing work structured in form, the modelās reasoning steps frequently remain superficial and exhibit inconsistent strategic coherence. In summary, existing approaches suffer from three key limitations. (1) Incomplete Empathy Strategy Coverage. Prior methods lack a com- prehensive strategy system covering diverse emo- tional states and higher-order cognitive strategies, limiting principled decision-making. (2) Lack of Task-Aligned Multi-Stage Reasoning. CoT-based methods do not explicitly model dialogue as a task- specific, multi-stage reasoning process, constrain- ing deep deliberation and response quality. (3) In- sufficient Strategy-Aware Supervision. Training data provide few high-quality annotations aligned with empathetic strategies and reasoning. To address these issues, we propose STRIDE- ED, a framework that proceeds by first establishing a comprehensive strategy system, which then en- ables task-aligned reasoning through a dedicated data pipeline. At its core, we construct a uni- fied strategy system covering positive, neutral, and negative emotions to guide response generation. Leveraging this system, we automatically annotate the EMPATHETICDIALOGUES dataset (Rashkin et al., 2019) with strategy types and rationales using authoritative LLMs. Subsequently, a rigorous data refinement process employs multi-LLM evaluation with consistency-weighted scoring and strategy- aware sampling to curate high-quality training sub- sets. Finally, the model is optimized through a two-stage training paradigm that combines super- vised fine-tuning with multi-objective reinforce- ment learning, ensuring alignment with strategic, emotional, and structural correctness. Our contri- butions can be summarized as follows: ā¢We propose an interpretable framework for empathetic dialogue, STRIDE-ED, featuring a comprehensive empathy strategy system and stepwise decision-making. ā¢To support effective training, a strategy-aware data refinement pipeline is constructed, com- bining LLM-based annotation, multi-model weighted evaluation, and dynamic sampling to regulate strategy distribution and difficulty. ā¢A two-stage training paradigm is introduced, in which supervised fine-tuning establishes reasoning, followed by reinforcement learn- ing to improve emotional alignment, strategy execution, and response consistency. ā¢Extensive experiments show that our frame- work generalizes across diverse open-source LLMs and achieves superior performance on both automatic and human evaluations. 2 Related Works 2.1 Implicit Knowledge-Driven Empathy Early empathetic dialogue models lacked sufficient prior knowledge, which limited their ability to understand usersā emotions and contexts. To ad- dress this issue, some studies incorporate external knowledge into dialogue modeling to implicitly en- hance reasoning and decision-making (Lee et al., 2022). For example, Ghosal et al. (2020) lever- aged structured commonsense knowledge from ATOMIC (Sap et al., 2019), such as mental states and causal relations, to model interlocutor in- teractions for improved emotion understanding. Zhong et al. (2021) integrated commonsense-aware emotional latent concepts to generate emotion- ally appropriate responses, while Sabour et al. (2022) inferred usersā situational contexts through commonsense-based reasoning.Further, Cai et al. (2023) introduced an adaptive commonsense knowledge selection mechanism to refine contex- tual cognition, and Qiao et al. (2025) constructed multi-hop reasoning graphs to incorporate external knowledge during response generation. However, these methods did not explicitly model reasoning or strategy-guided decision-making. 2.2Explicit Reasoning with CoT for Empathy LLMs possessed extensive knowledge reserves, and CoT prompting (Wei et al., 2022) enabled step- wise reasoning, facilitating more explicit decision- making in empathetic dialogue. Building on this paradigm, Tu et al. (2022) inferred fine-grained user emotions and employed a mixed strategy mechanism for response generation, while Chen and Liu (2023) dynamically generated counseling strategies via zero-shot prompting to guide per- sonalized responses. Subsequent work further en- hanced empathy through data and structural de- signs: Chen et al. (2023) fine-tuned LLMs on consultant-style multi-turn dialogues, and Ye et al. (2025) adopted a strategy-enhanced role-playing framework with multiple interacting roles to gen- erate diverse training data. Finally, Zhang et al. (2025) introduced an intention-centered framework that mapped inferred supporter intentions to sup- port strategies using a chain-of-thought mechanism. Despite improved interpretability, these CoT-based approaches lacked comprehensive strategy cover- age and did not explicitly model the full, structured reasoning process underlying human empathetic decision-making. 3 Methodology STRIDE-ED is a general-purpose framework for empathetic dialogue applicable to diverse open- source LLMs. It implements a comprehensive em- pathy strategy system and a task-aligned, multi-step CoT paradigm to model dialogue as a progressive cognitive and decision-making process. At the data and training levels, STRIDE-ED integrates LLM- based automatic annotation, strategy-aware data re- finement, and two-stage optimization. An overview is shown in Figure 2. 3.1 Task Formulation Empathetic dialogue consists of a multi-turn in- teraction between a user and a conversational agent. We denote the dialogue history asC = u 1 ,u 2 ,...,u tā1 , whereu i represents the utter- ance at thei-th turn. Each utteranceu i is a se- quence of tokens, i.e.,u i = (w i,1 ,w i,2 ,...,w i,n i ). The goal of empathetic dialogue is to generate an empathetic responseu t while appropriately recog- nizing the userās emotional statee. In this work, we introduce auxiliary objectives and model empa- thetic dialogue as a sequential generation process. Specifically, conditioned onC, the model gener- ates a dialogue scenario summarysum, infers the target emotional statee, determines an empathy strategystraalong with its execution actionsacts, and finally produces the empathetic responseu t , corresponding to modeling the conditional distribu- tion P(sum,e, stra, acts,u t |C). 3.2 Empathy Strategy System In empathetic dialogue modeling, response strate- gies form a crucial intermediate stage between emo- tion understanding and response generation. Rather than generating surface-level replies directly, ef- fective models must explicitly decide how to re- spond based on the userās emotional state and dia- logue context. Liu et al. (2021) proposed a widely adopted taxonomy of eight empathy strategies, pri- marily applied in counseling-oriented settings that focus on negative emotions. Higher-order cognitive strategies, however, remain underexplored, limit- ing the effectiveness of current systems in complex dialogues requiring sophisticated reasoning. Motivated by these observations, we first analyze the emotional distributions in the EMPATHETIC- DIALOGUES dataset and conduct a fine-grained examination of response content. Based on these analyses, we expand the original empathy strategy set to better accommodate positive, neutral, and negative emotional contexts, and assign a three- level difficulty rating (IāI) to each strategy to re- flect the cognitive complexity involved in its appli- cation. In particular, the original Question strategy is further subdivided into three distinct Exploring- type strategies to capture more nuanced cognitive and emotional interactions. The structured strategy system guides both annotation and model training (full details are provided in Appendix A). 3.3 Stepwise Cognitive CoT Design Inspired by insights from cognitive psychol- ogy (Lee et al., 2023), we model the human thought process during empathetic dialogue as a stepwise, incremental deliberation. Upon receiving a speakerās utterance, individuals first infer miss- ing information or make educated guesses about the described situation based on prior knowledge and contextual understanding, forming a high-level situational representation. This situational com- prehension allows them to adopt the speakerās per- spective and facilitates accurate recognition of the speakerās emotional state. Next, humans deliberate on which response strategy to employ, guided by the speakerās emotional state and contextual cues. Because strategies represent generalized methods, their concrete execution may differ across scenar- ios. To account for this variability, we introduce an ED-CSA-all LLM-Based Automated Data AnnotationData CurationTwo-Stage Training Pipeline Role: You are an expert in empathetic dialogue and strategy analysis. Task: This is an empathetic conversation, please read the context and focus on the listenerās last reply. Now, suppose you are the listener mentioned aboveāplease complete the following tasks: ... Role: You are an expert in empathetic dialogue and strategy analysis. Task: This is an empathetic conversation, please read the context and focus on the listenerās last reply. Now, suppose you are the listener mentioned aboveāplease complete the following tasks: ... Role: You are an evaluator. Task: Your task is to assess whether the reasoning process is reasonable and accurate in supporting the target reply, given the dialogue. Please output a single integer score from 1 to 5 ... Role: You are an evaluator. Task: Your task is to assess whether the reasoning process is reasonable and accurate in supporting the target reply, given the dialogue. Please output a single integer score from 1 to 5 ... Consistency-Based Scoring and Ranking Strategy-Aware Sampling Scheme ED-CSA-12kED-CSA-12k ScenarioScenarioStrategyStrategyActions DeepSeek-R1-671bDeepSeek-R1-671b ED-Original ContextEmotionResponseContextEmotionResponse ED-Original ContextEmotionResponse Context Strategy Scenario Actions Multi-Judge ScoringRanking and Selection ED-CSA-5kED-CSA-5k Sampling Scale Determination Strategy Distribution Scale Determination Strategy Distribution Stage 1 Supervised Fine-Tuning Stage 2 Proximal Policy Optimization R- CalculationR- CalculationR-strategyR-strategyR-strategyR-emotionR-emotionR-formatR-format Stage 2 Proximal Policy Optimization R- CalculationR-strategyR-emotionR-format LLMs Ā· Ā· Ā· LLMs Ā· Ā· Ā· <think> <Summary> Speaker planned to ... <Emotion> Lonely <Strategy> Cognitive Reframing: I would apply Cognitive Reframing by... </think> <answer> If the movie is worth it you will never miss them. </answer> <think> <Summary> Speaker planned to ... <Emotion> Lonely <Strategy> Cognitive Reframing: I would apply Cognitive Reframing by... </think> <answer> If the movie is worth it you will never miss them. </answer> STRIDE-ED-DeepSeek-7B Figure 2: The architecture of the STRIDE-ED framework, illustrating the complete pipeline from data preparation and refinement to model training. action inference stage that bridges strategy selec- tion and the generation of the final response. This stepwise CoT design captures the struc- tured cognitive progression in human empathetic reasoning, providing a principled framework for multi-stage, strategy-aware response generation. Implementation leverages structured intermedi- ate tagsāsuch as<Context>,<Emotion>, and <Strategy>āto guide the modelās internal reason- ing, culminating in the final response. 3.4 LLM-Based Automated Data Annotation Building upon the aforementioned theoretical framework, we address the absence of explicit an- notations required for model training. Specifically, the adopted dataset lacks annotations for dialogue scenario summaries, empathy strategy types, and the concrete actions used to implement each strat- egy. To this end, we employ DeepSeek-R1 1 , a large language model with strong reasoning capa- bilities, as an automated annotation expert. Care- fully designed annotation prompts are used to gen- erate structured labels for each dialogue instance, as shown in Appendix B. The annotated dataset re- sulting from this process is denoted as ED-CSA-all. 3.5 Consistency-Based Scoring and Ranking To ensure the reliability of the automatically anno- tated data while mitigating the high cost of large- scale manual verification, we employ a multi-judge 1 https://huggingface.co/deepseek-ai/DeepSeek-R1 evaluation mechanism based on LLMs. Specifi- cally, three representative models (DeepSeek-R1, Qwen3 2 , and Llama-3.1 3 ) are selected as indepen- dent expert evaluators. Using carefully designed evaluation prompts, as shown in Appendix B, each model assesses every example in the ED-CSA-all dataset and assigns a quality score ranging from 1 to 5, where higher scores indicate stronger semantic coherence and alignment among the annotated sce- nario summary, selected strategy, inferred action, and the corresponding dialogue context. Models are restricted to generate integer scores only. We apply a reliability-weighted multi-judge ag- gregation framework to fairly integrate scores from multiple LLM evaluators.LetM = m 1 ,m 2 ,m 3 denote the set of evaluators, cor- responding to DeepSeek-R1, Qwen, and LLaMA. For eachm i ā M, lets i denote its score vector over the dataset. The reliability of each evalua- tor is estimated as the average Spearman correla- tion (Spearman, 1904) with the others: Ļ i = 1 2 X jĢø=i Spearman(s i ,s j ), āiāM. (1) The final quality score for a samplexis com- puted as a reliability-weighted sum of evaluator 2 https://huggingface.co/Qwen/Qwen3-8B 3 https://huggingface.co/meta-llama/Llama-3.1-8B- Instruct I I I Positive Reinforcement (5.5%) Exploring Actions and Intentions (23.3%) Self-disclosure (18.7%) Affirmation and Reassurance (21.7%) Reflection of Feelings (5.1%) Cognitive Reframing (5.8%) Exploring Thoughts and Cognition (6.9%) Providing Suggestions (6.8%) Neutral Validation (0.5%) Information (1.1%) Others (0.4%) Restatement or Paraphrasing (1.7%) Gratitude Prompting (0.1%) Exploring Feelings and Emotions (2.4%) Figure 3: Distribution of Strategy Types in ED-CSA-all. scores with an agreement-based penalty: S(x) = |M| X i=1 Ļ i P |M| k=1 Ļ k s i (x)ā Ī»Ļ(x),(2) wheres i (x)denotes the score assigned to samplex by evaluatorm i , andĻ(x)represents the standard deviation of the scoress i (x) |M| i=1 across evalua- tors, measuring inter-rater disagreement. The hy- perparameterĪ»is set to0.1in our experiments. This formulation favors samples with both high weighted scores and strong evaluator consensus. Based on the resulting score ranking, we select the top 12k samples as the candidate pool for sub- sequent sampling, denoted as ED-CSA-12k. 3.6 Strategy-Aware Sampling Scheme We further emphasize that strategy usage should re- flect both distributional characteristics and contex- tual sensitivity rather than being applied uniformly. The distribution of strategy types in ED-CSA-all is illustrated in Figure 3. Data sorting and filtering inherently reshape the empirical strategy distribu- tion, and directly training on such processed data may distort natural response patterns. Simultane- ously, the model should learn to engage in selective strategy reasoning based on the complexity of the dialogue context, reserving explicit strategy and ac- tion deliberation for scenarios that demand higher cognitive effort. Based on the above considerations, we design a strategy-aware sampling scheme in which the target sampling distribution is jointly determined by the empirical frequency of each strategy and its associ- ated difficulty level. LetS =s 1 ,...,s |S| denote the set of empathy strategies. We first compute the empirical frequency of each strategys i in the full annotated dataset ED-CSA-all, denoted asa i . Next, we assign a difficulty weightd i to each strategy according to its difficulty level, where higher-level strategies receive larger weights. We then compute a weighted score for each strategy by combining its empirical frequency with its assigned difficulty level, and normalize these scores to obtain the tar- get sampling proportions: p i = a i Ā· d i P |S| j=1 a j Ā· d j .(3) This distribution favors the retention of infre- quent yet cognitively demanding strategies, while preventing low-difficulty strategies from dispro- portionately dominating the sampled data. Subse- quently, given the ranked candidate pool ED-CSA- 12k, we apply a binary search procedure to deter- mine the maximum subset size whose empirical strategy distribution can satisfy the target propor- tionsp i . Based on this subset size and the target distribution, we compute the required number of samples for each strategy and perform stratified random sampling to construct the refined training set ED-CSA-5k. 3.7 Model Training After constructing the refined training data, we pro- ceed to model training using a two-stage optimiza- tion paradigm. Supervised Fine-Tuning In the first stage, we construct the supervised training set by combining the curated subset ED-CSA-5k with the remaining samples from ED-CSA-all, for which strategy an- notations are removed. This design ensures that only the curated subset provides explicit strategy guidance, while the remaining samples contribute to general response generation without strategy su- pervision. The base LLMMis then fine-tuned to sequentially generate both the intermediate reason- ing components and the final response, yielding the model denoted as STRIDE-ED-M-SFT. The su- pervised training objective minimizes the negative log-likelihood over the structured output sequence: L SFT =ā 1 N N X n=1 logp Īø (y |C),(4) whereyrepresents the target output sequence con- ditioned on the dialogue historyC. Reinforcement Learning In the second stage, we further refine the SFT model via reinforcement learning on ED-CSA-12k. Beginning with STRIDE- ED-M-SFT, we apply Proximal Policy Optimiza- tion (PPO) to optimize the model towards outputs that better adhere to the desired format, emotional alignment, and strategy execution. The reward function is defined as follows: r format = ( 1, if format is correct 0, otherwise (5) r emotion = ( 1, if emotion is correct 0, otherwise (6) r strategy = ( 0, if strategy is incorrect 1, otherwise (7) R = r format Ā· 1 + r emotion + r strategy (8) wherer format ,r emotion , andr strategy denote the indi- vidual rewards for format, emotion, and strategy, respectively. The final rewardRemphasizes re- sponses with correct format while simultaneously incentivizing emotional and strategic correctness. Ultimately, we obtain the model STRIDE-ED-M. 4 Experiments 4.1 Datasets We adopt the EMPATHETICDIALOGUES dataset (Rashkin et al., 2019) for our experiments. Devel- oped by Facebook AI Research, this dataset is a large-scale multi-turn dialogue corpus dedicated to enhancing the empathetic capabilities of conversa- tional systems. It comprises approximately 25,000 dialogue scenarios grounded in specific emotional contexts, covering 32 emotion labels. In the dataset, participants are assigned the roles of speaker and listener to simulate human interactions around emo- tional experiences. 4.2 Baselines To ensure a comprehensive evaluation, we select baseline models spanning three major research tra- jectories in empathetic dialogue generation: (1) Traditional Architectures that focus on neural net- work design for emotional modeling, including Transformer (Vaswani et al., 2017), MoEL (Lin et al., 2019) and EmpDG (Li et al., 2020). (2) External-Knowledge-Enhanced Methods that in- corporate commonsense or causal graphs to en- rich responses, such as KEMP (Li et al., 2022), DCKS (Cai et al., 2023), E-CORE (Fu et al., 2023), and Emp-USIR (Jiang et al., 2023). (3) Reflective Decision-Integration Approaches that explicitly model internal cognitive processes (e.g., reasoning, reflection) for strategic response generation, repre- sented by CAB (Gao et al., 2023), IAMM (Yang et al., 2024), and ReflectDiffu (Yuan et al., 2025). 4.3 Implementation Details We implement all LLM training using the verl 4 framework, with DeepSeek-7B-Chat selected as the primary model. 5 During the supervised fine- tuning stage, the initial learning rate is set to1Ć 10 ā4 , the batch size is 16, and the maximum input sequence length is capped at 2048 tokens. In the reinforcement learning stage, the total batch size is increased to 128, with the maximum generation length limited to 1024 tokens. We follow a fixed 8:1:1 split for training, validation, and testing. All experiments are conducted on four NVIDIA A800 GPUs, each equipped with 80 GB of memory. 4.4 Evaluation Metrics To comprehensively evaluate STRIDE-ED, we con- duct both automatic and human assessments. Automatic Evaluation We evaluate response quality using Perplexity (PPL), BLEU (B-n), Dis- tinct (D-n), and emotion accuracy (Accemo). PPL measures fluency, with lower values indicating greater coherence. BLEU captures relevance to reference responses via n-gram overlap, while Dis- tinct reflects lexical diversity. Accemomeasures alignment of predicted and ground truth emotions. Human Evaluation We perform A/B testing to assess Empathy, Relevance, and Fluency. A panel of three annotators is employed, and 1000 dialogue turns generated by STRIDE-ED and baselines are selected for comparison, ensuring the consistency and interpretability of the evaluation outcomes. 5 Results and Discussion Automatic Evaluation Results As shown in Ta- ble 1, the peak performance of our method across multiple runs consistently outperforms all base- line models across automatic evaluation metrics, demonstrating clear advantages in response rel- evance, emotional controllability, diversity, and fluency. Compared with the strongest baseline, ReflectDiffu, our method achieves consistent im- provements across all BLEU metrics, with rela- tive gains of 4.5% on BLEU-1, 8.7% on BLEU-2, 4 https://github.com/volcengine/verl 5 https://huggingface.co/deepseek-ai/deepseek-llm-7b- chat ModelsB-1āB-2āB-3āB-4āAcc emo āD-1āD-2āPPLā Transformer18.078.344.572.86ā0.361.3537.62 MoEL18.028.674.352.7331.020.431.7636.81 EmpDG19.969.114.742.8031.650.461.9937.43 KEMP18.078.304.372.6536.400.552.2936.89 DCKS21.7310.626.244.09 ā2.199.6116.08 E-CORE19.775.653.282.11ā0.683.3833.04 Emp-USIR20.119.865.723.73ā0.663.1035.29 CAB20.239.394.963.0140.520.892.9535.06 IAMM19.518.744.863.3243.720.883.0525.94 ReflectDiffu23.59 11.255.353.6248.760.984.3524.56 Ours24.6612.237.274.7657.572.5913.689.26 Table 1: Automatic evaluation results on the EMPATHETICDIALOGUES dataset. Boldface and underline indicate the best and second-best values, respectively. "Ours" refers to STRIDE-ED-DeepSeek-7B. ComparisonAspectsWinLoseTie Ours vs. MOEL Emp.46.123.630.3 Rel.39.822.437.8 Flu.33.515.650.9 Ours vs. EmpDG Emp.51.618.529.9 Rel.50.315.334.4 Flu.35.113.551.4 Ours vs. CAB Emp.53.223.223.6 Rel.56.621.422.0 Flu.31.511.656.9 Table 2: Human A/B evaluation results. 35.9% on BLEU-3, and 31.5% on BLEU-4, indi- cating stronger alignment with empathetic ground- truth responses at both surface and higher-order n- gram levels. In terms of emotional controllability, our method improves emotion accuracy by 18.1% over ReflectDiffu inAcc emo , indicating more re- liable emotion recognition. Our method also ex- hibits markedly enhanced lexical diversity, achiev- ing high Distinct-1 and Distinct-2 scores, which indicates its ability to generate richer and less repet- itive responses rather than relying on fixed empa- thetic templates. Furthermore, our method achieves low perplexity, reflecting improved fluency and overall generation confidence. Overall, these re- sults indicate that our method not only surpasses prior empathetic dialogue models but also achieves a more balanced trade-off among relevance, diver- sity, and fluency, validating the effectiveness of its structured reasoning and training framework. Human Evaluation Results Table 2 presents hu- man A/B testing results for STRIDE-ED and sev- eral baseline models on empathy (Emp.), relevance (Rel.), and fluency (Flu.). The results highlight the performance characteristics of each model in generating empathetic, coherent, and contextually relevant responses. ModelsB-1āAcc emo ā D-2āPPLā Ours24.6657.5713.689.26 w/o e23.58-13.66 10.47 w/o sum22.9154.1413.427.78 w/o stra22.4454.5815.096.98 w/o CoT22.55-13.268.64 w/o R.&S.22.2255.9315.25 11.86 w/o S.23.4456.0514.409.89 w/o PPO23.5254.4814.762.03 Table 3: Evaluation results of the ablation study.R.&S. refer to the two modules: Consistency-Based Scoring and Ranking, and Strategy-Aware Sampling Scheme. Ablation Study. Table 3 reports the ablation results of STRIDE-ED, covering both the stepwise cogni- tive reasoning components, data curation strategies, and PPO training. Overall, removing any compo- nent leads to performance degradation in one or more dimensions, confirming each module con- tributes non-trivially to the final model behavior. Stepwise Cognitive CoT Design. Removing emo- tion reasoning (w/oe) causes a clear drop in BLEU- 1 and an increase in PPL, indicating that explicit emotion modeling not only improves emotional alignment but also stabilizes generation quality. Eliminating scenario summarization (w/osum) re- sults in further degradation in BLEU-1 and emotion accuracy, suggesting that concise situation abstrac- tion is crucial for grounding subsequent emotional and strategic reasoning. When strategy reasoning is removed (w/ostra), emotion accuracy drops and fluency degrades, while diversity increases. This suggests that without strategy constraints, the model generates more varied but less controlled responses. Overall, strategy reasoning serves as a critical coordinating component in the framework, helping balance emotional controllability, coher- ence, and expressive diversity. Notably, removing structured reasoning (w/o CoT) leads to a drop in BLEU-1 while slightly reducing perplexity, sug- gesting that although responses become marginally easier to generate, they lose contextual relevance and cognitive grounding. This underscores the role of the stepwise cognitive CoT in enhancing re- sponse quality through structured reasoning rather than merely improving fluency. Data Scoring and Sampling. Removing both consistency-based scoring and strategy-aware sam- pling (w/oR.&S.) leads to the most severe degra- dation in BLEU-1, emotion accuracy, and perplex- ity, underscoring the role of systematic data fil- tering and distribution-aware sampling. Notably, while Dist-2 improves slightly, this likely reflects increased superficial diversity at the cost of co- herence and relevance, as indicated by declines in other metrics. When strategy-aware sampling is replaced with random sampling at the same scale (w/oS.), performance still drops compared to the full model. This confirms a balanced, difficulty- aware strategy distribution is essential, and naive random sampling can distort strategic learning. PPO Training. Removing the PPO stage results in a pronounced drop in both perplexity and emo- tion accuracy, reflecting a regression to generic and uncalibrated outputs. This highlights that PPO is essential for aligning the model along key di- mensions, including proper formatting, strategy ad- herence, and emotional fidelity, extending beyond mere instruction following. Training Set Size Analysis Table 4 shows the re- sults of training data size analysis on ED-CSA-all. Model competence remains stable until data is re- duced to 1/8, beyond which performance collapses sharply at 1/16āindicating severe overfitting and loss of task capability. The zero-data setup shows that while the pretrained model retains strong gen- erative fluency, it cannot adapt to the task without in-domain training. Overall, sufficient data is essen- tial for effective adaptation, with a clear minimum required to maintain robust performance. ModelsB-1āAcc emo āD-2āPPLā Ours24.6657.5713.689.26 1/2 Data23.4356.3515.07 12.75 1/4 Data23.3755.1214.568.80 1/8 Data 23.2851.2914.509.70 1/16 Data3.3245.533.386.21 Zero Data14.86ā21.491.09 Qwen3-0.6B21.2951.0010.469.79 Qwen3-4B20.4956.3512.51 10.20 Qwen3-4B-In.22.7655.9314.327.87 LLama3.2-3B22.7757.0614.49 11.11 GLM-Z1-9B22.1456.9814.34 10.51 Table 4: Evaluation results for training set size and backbone model analyses. The abbreviation "Qwen3- 4B-In." denotes Qwen3-4B-Instruct. Backbone Analysis Our method exhibits broad applicability and competitive performance across diverse LLM architectures and scales, as shown in Table 4. In the Qwen family, emotion accuracy improves with model size, whereas BLEU-1 does not scale monotonicallyāshowing that compact models can match or surpass larger ones in fluency with our approach. Across Qwen, LLama, and GLM families, the framework delivers consistently strong results: LLama3.2-3B leads in emotion accu- racy and diversity, while Qwen3-4B-Instruct excels in BLEU-1 and fluency. Details of the backbone LLMs are provided in Appendix C. These findings confirm that our framework is architecture-agnostic and parameter-efficient, delivering robust perfor- mance across different LLMs and scales. Case Study Representative case studies are pre- sented to demonstrate our framework, with details provided in Appendix D. 6 Conclusion In this paper, we present STRIDE-ED, a struc- tured empathetic dialogue framework that explic- itly models stepwise cognitive reasoning over situ- ation, emotion, and strategy. To support effective learning, we introduce a consistency-based scoring mechanism and a strategy-aware sampling scheme to construct supervision. Extensive experiments demonstrate that STRIDE-ED achieves improve- ments in empathetic controllability, response rele- vance, and lexical diversity across multiple settings. These findings suggest that incorporating explicit cognitive structure and data refinement is a promis- ing direction for empathetic dialogue systems. Limitations Our method currently lacks rigorous statistical val- idation in terms of data filtering and sampling. We plan to further investigate the underlying statis- tical principles and mechanisms to improve the soundness and rigor of the method design. In ad- dition, our experiments indicate that training with DeepSeek-7B-Chat yields the best results; how- ever, we have not yet examined whether this is related to the use of DeepSeek-R1 for data annota- tion. Furthermore, due to the high computational and time costs, we did not perform extensive hy- perparameter tuning on other models. Future work will explore the performance limits of individual backbone models more thoroughly. Ethical considerations LLMs are used for automatic annotation under prompts enforcing ethical constraints, and they are also applied to the grammatical checking and pol- ishing of texts. Human evaluators are informed that the evaluated content may contain negative emotions and are compensated fairly for their con- tributions. No personal or sensitive information is involved, and all experiments are conducted on publicly available datasets. References C Daniel Batson. 2009. These things called empathy: Eight related but distinct phenomena. The Social Neuroscience of Empathy, pages 3ā16. Hua Cai, Xuli Shen, Qing Xu, Weilin Shen, Xiaomei Wang, Weifeng Ge, Xiaoqing Zheng, and Xiangyang Xue. 2023. Improving empathetic dialogue genera- tion by dynamically infusing commonsense knowl- edge. In Findings of the Association for Computa- tional Linguistics: ACL 2023, pages 7858ā7873. Qi Chen and Dexi Liu. 2023. Dynamic strategy chain: Dynamic zero-shot cot for long mental health support generation. arXiv preprint arXiv:2308.10444. Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. Soulchat: Improving llmsā empathy, listening, and comfort abilities through fine-tuning with multi-turn empathy conversations. In Findings of the Associa- tion for Computational Linguistics: EMNLP 2023, pages 1170ā1183. Mark H Davis. 1983. Measuring individual differences in empathy: evidence for a multidimensional ap- proach. Journal of personality and social psychology, 44(1):113ā126. Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, and 1 others. 2024. The llama 3 herd of models. arXiv e-prints, pages arXivā2407. Fengyi Fu, Lei Zhang, Quan Wang, and Zhendong Mao. 2023. E-core: Emotion correlation enhanced em- pathetic dialogue generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 10568ā10586. Pan Gao, Donghong Han, Rui Zhou, Xuejiao Zhang, and Zikun Wang. 2023. Cab: empathetic dialogue generation with cognition, affection and behavior. In International Conference on Database Systems for Advanced Applications, pages 597ā606. Springer. Deepanway Ghosal, Navonil Majumder, Alexander Gel- bukh, Rada Mihalcea, and Soujanya Poria. 2020. Cosmic: Commonsense knowledge for emotion iden- tification in conversations. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2020, pages 2470ā2481. He Hu, Yucheng Zhou, Juzheng Si, Qianning Wang, Hengheng Zhang, Fuji Ren, Fei Ma, and Laizhong Cui. 2025. Beyond empathy: Integrating diagnostic and therapeutic reasoning with large language models for mental health counseling. arXiv e-prints, pages arXivā2505. Liting Jiang, Di Wu, Yanbing Li, and Wushouer Silamu. 2023. Emp-usir: A unidirectional synchronous in- teractive reasoning model for empathetic dialogue. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1ā9. IEEE. Jing Yang Lee, Kong Aik Lee, and Woon Seng Gan. 2022.Improving contextual coherence in varia- tional personalized and empathetic dialogue agents. In ICASSP 2022-2022 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 7052ā7056. IEEE. Yoon Kyung Lee, Inju Lee, Minjung Shin, Seoyeon Bae, and Sowon Hahn. 2023. Chain of empathy: En- hancing empathetic response of large language mod- els based on psychotherapy models. arXiv preprint arXiv:2311.04915. Qintong Li, Hongshen Chen, Zhaochun Ren, Pengjie Ren, Zhaopeng Tu, and Zhumin Chen. 2020. Empdg: Multi-resolution interactive empathetic dialogue gen- eration. In Proceedings of the 28th international conference on computational linguistics, pages 4454ā 4466. Qintong Li, Piji Li, Zhaochun Ren, Pengjie Ren, and Zhumin Chen. 2022. Knowledge bridging for em- pathetic dialogue generation. In Proceedings of the AAAI conference on artificial intelligence, volume 36, pages 10993ā11001. Zhaojiang Lin, Andrea Madotto, Jamin Shin, Peng Xu, and Pascale Fung. 2019. Moel: Mixture of empa- thetic listeners. In Proceedings of the 2019 Confer- ence on Empirical Methods in Natural Language Pro- cessing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 121ā132. Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, and 1 others. 2024. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437. Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. 2021. Towards emotional support dialog systems. In Proceedings of the 59th Annual Meet- ing of the Association for Computational Linguistics and the 11th International Joint Conference on Natu- ral Language Processing (Volume 1: Long Papers), pages 3469ā3483. Ye Liu, Wolfgang Maier, Wolfgang Minker, and Stefan Ultes. 2022. Empathetic dialogue generation with pre-trained roberta-gpt2 and external knowledge. In Conversational AI for Natural Human-Centric In- teraction: 12th International Workshop on Spoken Dialogue System Technology, IWSDS 2021, Singa- pore, pages 67ā81. Springer. Baiyou Qiao, Yuekai Zhang, Pan Gao, Xinchi Li, Shuo Wang, and Donghong Han. 2025. Multi-perspective empathy modeling for empathetic dialogue genera- tion. Knowledge-Based Systems, 314:113191. Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. 2019. Towards empathetic open- domain conversation models: A new benchmark and dataset. In Proceedings of the 57th annual meeting of the association for computational linguistics, pages 5370ā5381. Sahand Sabour, Chujie Zheng, and Minlie Huang. 2022. Cem: Commonsense-aware empathetic response gen- eration. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11229ā 11237. Maarten Sap, Ronan Le Bras, Emily Allaway, Chan- dra Bhagavatula, Nicholas Lourie, Hannah Rashkin, Brendan Roof, Noah A Smith, and Yejin Choi. 2019. Atomic: An atlas of machine commonsense for if- then reasoning. In Proceedings of the AAAI con- ference on artificial intelligence, volume 33, pages 3027ā3035. C Spearman. 1904. The proof and measurement of as- sociation between two things. The American Journal of Psychology, 15(1):72ā101. Quan Tu, Yanran Li, Jianwei Cui, Bin Wang, Ji-Rong Wen, and Rui Yan. 2022. Misc: A mixed strategy- aware model integrating comet for emotional support conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 308ā319. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Åukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022. Chain-of-thought prompting elic- its reasoning in large language models. Advances in neural information processing systems, 35:24824ā 24837. Zhichao Xu and Jiepu Jiang. 2024. Multi-dimensional evaluation of empathetic dialogue responses. In Find- ings of the Association for Computational Linguistics: EMNLP 2024, pages 2066ā2087. An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025.Qwen3 technical report.arXiv preprint arXiv:2505.09388. Zhou Yang, Zhaochun Ren, Wang Yufeng, Haizhou Sun, Chao Chen, Xiaofei Zhu, and Xiangwen Liao. 2024. An iterative associative memory model for em- pathetic response generation. In Proceedings of the 62nd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 3081ā3092. Jing Ye, Lu Xiang, Yaping Zhang, and Chengqing Zong. 2025. Sweetiechat: A strategy-enhanced role- playing framework for diverse scenarios handling emotional support agent. In Proceedings of the 31st International Conference on Computational Linguis- tics, pages 4646ā4669. Jiahao Yuan, Zixiang Di, Zhiqing Cui, Guisong Yang, and Usman Naseem. 2025. Reflectdiffu: Reflect be- tween emotion-intent contagion and mimicry for em- pathetic response generation via a rl-diffusion frame- work. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 25435ā25449. Tenggan Zhang, Xinjie Zhang, Jinming Zhao, Li Zhou, and Qin Jin. 2024. Escot: Towards interpretable emotional support dialogue systems. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13395ā13412. Xinjie Zhang, Wenxuan Wang, and Qin Jin. 2025. In- tentionesc: An intention-centered framework for en- hancing emotional support in dialogue systems. In Findings of the Association for Computational Lin- guistics: ACL 2025, pages 26494ā26516, Vienna, Austria. Association for Computational Linguistics. Peixiang Zhong, Di Wang, Pengfei Li, Chen Zhang, Hao Wang, and Chunyan Miao. 2021.Care: commonsense-aware emotional response generation with latent concepts. In Proceedings of the AAAI Con- ference on Artificial Intelligence, volume 35, pages 14577ā14585. A Strategy System Details The comprehensive strategy system underpinning the STRIDE-ED framework is detailed in Table 5. It encompasses 14 distinct strategies categorized into three tiers of cognitive and implementational Difficulty (I, I, and I). This taxonomy is con- structed to guide the modelās reasoning chain across the full emotional spectrum, ranging from fundamental responses like Restatement (Difficulty I) to advanced cognitive interventions such as Cog- nitive Reframing (Difficulty I). The tiered struc- ture ensures that the framework can adaptively en- gage in appropriate strategic planning, mirroring the nuanced decision-making process of human empathetic dialogue. B Prompts In the research on conversational strategy analysis and reasoning quality evaluation of large language models, standardized prompts are the key to ensur- ing the consistency and reliability of task results. Below, we will introduce the design concepts and principles of two core prompts. For the Data Anno- tation Prompt, it achieves high-quality strategy clas- sification annotation and clarifies the reasons for strategy selection through professional role guid- ance, closed-set strategy selection, and standard- ized output. The second one is the Consistency- Based Scoring Prompt, which performs minimal quantitative scoring from a neutral evaluation per- spective and completes an objective assessment of the rationality of model reasoning by invoking three different models. Data Annotation Prompt Role: You are an expert in empathetic dialogue and strategy analysis. Inputs: 1. Context: 2. Emotion: 3. Listenerās Response: Task: This is an empathetic conversation, please read the context and focus on the listenerās last reply. Now, suppose you are the listener mentioned aboveāplease complete the following tasks: <Summary>Brieflysummarizethe speakerās situation. Data Annotation Prompt (Continued) <Strategy> From the first-person perspec- tive, choose one strategy from the optional strategies and their interpretations below and explain how you would apply it, keeping the reasoning concise. Optional strategies: 1.Exploring Thoughts and Cognition: Probes the speakerās beliefs, interpretations, and thought processes. 2.Exploring Actions and Intentions: ... ... 14.Others: ... Output Format: <Summary> ... <Strategy> [one strategy], [reason and actions] Output Requirement: 1. Focus on the speakerās last utterance for the need. 2. Pick only strategies actually used in the listenerās response. 3. Be concise and precise. Consistency-Based Scoring Prompt Role: You are an evaluator. Inputs: 1.The dialogue context. 2.The target reply. 3.The generated reasoning process. Task: Your task is to assess whether the rea- soning process is reasonable and accurate in supporting the target reply, given the dia- logue. Please output a single integer score from 1 to 5 (1 = very poor, 5 = excellent). Output Requirement: Output only the in- teger score (1-5), without any explanation, extra text, punctuation, or formatting. C Experimental Details This appendix section provides an overview of the backbone LLMs used in our experiments, and in- formation regarding the involved existing packages is available in the code repository 6 . Qwen3-0.6B 7 is a 0.6B-parameter causal lan- guage model with 28 layers and a 32,768-token con- text. It supports thinking and non-thinking modes, 6 https://anonymous.4open.science/r/STRIDE-ED/ 7 https://huggingface.co/Qwen/Qwen3-0.6B StrategyDefinitionDifficulty 1Gratitude Prompting Encourages the speaker to notice and reflect on positive experiences or supportive aspects of their life, fostering positive emotional awareness. I 2Restatement or Paraphrasing Rephrases the speakerās main ideas to demonstrate un- derstanding and attentive listening. I 3Others Covers responses that do not clearly fit into any prede- fined strategy category. I 4Information Provides objective facts or relevant knowledge to help the speaker better understand their situation or make decisions. I 5Neutral Validation Affirms that neutral or low-intensity emotional states are normal and acceptable without encouraging stronger emotions. I 6Positive Reinforcement Highlights the speakerās strengths, efforts, or construc- tive behaviors to reinforce confidence and motivation. I 7 Exploring Actions and Intentions Uses targeted questions to clarify the speakerās actions, plans, and underlying intentions. I 8Self-disclosure Shares limited, relevant personal information to foster rapport and mutual understanding. I 9Affirmation and Reassurance Acknowledges the speakerās feelings and offers com- fort or emotional support. I 10Reflection of Feelings Identifies and articulates emotions that the speaker im- plies but does not explicitly express. I 11Cognitive Reframing Offers an alternative perspective on a difficult situation while respecting the speakerās original emotions. I 12 Exploring Feelings and Emotions Uses open-ended prompts to encourage deeper expres- sion of the speakerās emotional experience. I 13 Exploring Thoughts and Cognition Probes the speakerās beliefs, interpretations, and thought processes. I 14Providing Suggestions Offers practical and actionable recommendations tai- lored to the speakerās needs. I Table 5: The Structured Empathy Strategy System of STRIDE-ED. demonstrating strong reasoning, agent capabilities, and multilingual instruction following. Qwen3-4B 8 is a 4B-parameter causal language model with 36 layers and a a 32,768 natively and 131,072 tokens with YaRN context. It features en- hanced multilingual understanding, high-efficiency inference, and robust performance on Chinese and English tasks, excelling in factual accuracy, cre- ative generation, and industrial scenario adaptation. Qwen3-4B-Instruct 9 is a 4B-parameter causal language model with 36 layers. Optimized for instruction following, it supports zero-shot/few- shot learning, real-time dialogue interaction, and domain-specific task customization, demonstrating superior alignment with human intent and practical application scalability. LLama3.2-3B 10 is a 3B-parameter causal lan- guage model with 28 layers and a 8,192-token con- text. It focuses on lightweight deployment, deliver- ing balanced performance in reasoning, code gener- ation, and multilingual processing, with optimized efficiency for edge and low-resource environments. GLM-Z1-9B 11 is a 9B-parameter causal lan- guage model with 40 layers and a 30,000-token con- text. It integrates self-supervised pre-training with instruction fine-tuning, showcasing strong mathe- matical reasoning, long-text comprehension, and cross-modal fusion capabilities, tailored for com- plex industrial and academic reasoning tasks. D Case Study In this appendix, we present a set of representa- tive case studies covering higher-difficulty strate- gies, including Providing Suggestions, Exploring Thoughts and Cognition, Cognitive Reframing, etc., to illustrate the performance of our model. These examples are shown in Tables 6 - 11. Each case provides multi-dimensional core in- formation: dialogue context, high-level empathetic response strategy, model-generated response, and real-scenario reference result. Side-by-side pre- sentation of this information intuitively illustrates the modelās strategy selection logic and linguistic expression ability in complex conversations, and allows direct comparison between model outputs and real results to accurately evaluate the modelās performance in applying high-level empathetic re- sponse strategies. 8 https://huggingface.co/Qwen/Qwen3-4B 9 https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507 10 https://huggingface.co/meta-llama/Llama-3.2-3B 11 https://huggingface.co/zai-org/GLM-Z1-9B-0414 ContextSpeaker: I am going to a concert very soon. SummaryThe speaker is anticipating an upcoming concert they will at- tend. Emotion Excited StrategyExploring Actions and Inten- tions ActionsI chose Exploring Actions and Intentions because the speakerās excitement about the concert hints at unfolding plans, and ask- ing about details channels that energy into shared anticipation. Generated ResponseThatās awesome! Who are you going to see? Ground TruthCool! Who are you going to see? Table 6: Case Study 1 illustrating the use of the Cogni- tive Reframing strategy ContextSpeaker: Speaking of kids in school, I just found out my son got a B in his world history class at the end of the previous school year. I canāt believe it. Listener: Thatās great!! His- tory is so important I think and so few kids really grasp it well. Congrats to him for a job well done! Speaker: No no, see, it was the first grade below an A that heās ever made. I honestly thought I had the next Einstein on my hands. Now I guess Iāl have to settle for...above average, I guess. :( SummaryThe speakerās son achieved a B in history, his first grade below an A, which initially made the speaker hopeful but now feels disappointed. Emotion surprised StrategyCognitive Reframing ActionsI chose Cognitive Reframing be- cause the speakerās disappoint- ment over a single B clashed with sky-high expectations, and recasting the grade as proof of challenging coursework softens the sting and restores perspec- tive. Generated ResponseWell, thatās still a great grade! Iām sure heāl do better next year. Iām sure youāre proud of him! Ground TruthOh, I see what you are saying! Did he get upset about it? And you never know, he could turn it around next year and blow you away! Table 7: Case Study 2 illustrating the use of the Cogni- tive Reframing strategy ContextSpeaker: Every year when the kids go back to school, it brings back so many memories of when my kids were going to school. Listener: I know what you mean. I love to see all the pictures, and think back to how little mine used to be. Speaker: I would cry the first day of school every year, es- pecially the milestone years. Now my grandchildren are start- ing school. Time flies by too quickly SummaryThe speaker reflects on their childrenās school years, now ex- periencing the same bittersweet emotions as their grandchildren begin school, expressing how quickly time passes. EmotionSentimental StrategySelf-disclosure ActionsI chose Self-disclosure because the speakerās nostalgia over grandchildren starting school echoes my own disbelief at how fast kids grow, and sharing that sense of fleeting time affirms weāre in the same bittersweet boat. Generated ResponseI know. I have a 10 year old and a 13 year old, and I canāt believe how fast it has gone by. Iām sure you feel the same way. Ground TruthThey do seem to be grown be- fore you know it. Make you wonder how if goes so fast! Table 8: Case Study 3 illustrating the use of the Self- disclosure strategy ContextSpeaker: I am going to my local amusement park tomorrow and feeling a certain way about it! SummaryThe speaker is anticipating a visit to an amusement park and feels a mix of excitement and anxiety about the experience. Emotion Apprehensive StrategyExploring Thoughts and Cogni- tion ActionsI chose Exploring Thoughts and Cognition because the speakerās mixed excitement-and- apprehension about the amuse- ment park suggests hidden wor- ries, and probing the thoughts behind the nerves can turn anxi- ety into eager curiosity. Generated ResponseThat sounds like a lot of fun! Are you nervous? Ground TruthWhich way are you feeling? Are you a thrill seeker? Table 9: Case Study 4 illustrating the use of the Explor- ing Thoughts and Cognition strategy ContextSpeaker: Iām so happy summer classes are officially over and Iāl be able to relax for a week! Listener: Great what will you be doing during your time off ? Speaker: Iām not even sure what to do with all this time! I guess I should relax or have some leisure time, but everything in me says to try and be productive even on break. Itās that school mentality. SummaryThe speaker is conflicted about balancing relaxation and pro- ductivity during their week-long break, struggling with the linger- ing mindset of being productive despite wanting downtime. Emotion Joyful StrategyProviding Suggestions ActionsI chose Providing Suggestions because the speaker felt torn be- tween rest and productivity, and offering concrete ways to enjoy downtime eases the guilt while keeping the momentum alive. Generated ResponseI think you should just relax and enjoy the time off. You can al- ways get back to work after that. Ground TruthI agree with you,youāl feel good about yourself being productive ! Table 10: Case Study 5 illustrating the use of the Pro- viding Suggestions strategy. ContextSpeaker: The litter of puppies had been born! Listener: Nice, amazing to wit- ness birth. Speaker: All of the puppies sur- vived, truly a lucky amazing oc- casion! SummaryThe speaker is celebrating the successful birth of their pup- pies and expressing gratitude for their survival. Emotion joyful StrategyAffirmation and Reassurance ActionsI chose Affirmation and Re- assurance because the speaker is deeply joyful about their puppiesā birth and expressing thanks for their survival; this re- inforces their positive feelings and validates their relief by ac- knowledging the success and wishing them well. Generated ResponseThat is a great news, congrats. Ground TruthGood to hear, have fun with pup- pies! Table 11: Case Study 6 illustrating the use of the Affir- mation and Reassurance strategy