STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems

Hongru Ji; Yuyin Fan; Meng Zhao; Xianghua Li; Lianwei Wu; Chao Gao

cs.CLApr 8, 2026

STRIDE-ED: A Strategy-Grounded Stepwise Reasoning Framework for Empathetic Dialogue Systems

Authors: Hongru Ji, Yuyin Fan, Meng Zhao, Xianghua Li, Lianwei Wu, Chao Gao

Organizations: School of Artificial Intelligence, OPtics and ElectroNics (iOPEN), Northwestern Polytechnical University, China · School of Artificial Intelligence and Big Data, Henan University of Technology, China · School of Computer Science, Northwestern Polytechnical University, China

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. Our data and code are publicly available at https://github.com/jicoder-nwpu/STRIDE-ED.

Figures & tables

Figure 1: Illustration of the STRIDE-ED reasoning process. Given a user utterance, the framework performs scenario summarization, emotion recognition, strategy inference, and strategy-guided response generation.

Figure 2: The architecture of the STRIDE-ED framework, illustrating the complete pipeline from data preparation and refinement to model training.

Figure 3: Distribution of Strategy Types in ED-CSA-all .

Models	B-1 $\uparrow$	B-2 $\uparrow$	B-3 $\uparrow$	B-4 $\uparrow$	$\textbf{Acc}_{emo}\uparrow$	D-1 $\uparrow$	D-2 $\uparrow$	BARTScore $\uparrow$	PPL $\downarrow$
Transformer	18.07	8.34	4.57	2.86	–	0.36	1.35	-	37.62
MoEL	18.02	8.67	4.35	2.73	31.02	0.43	1.76	0.56	36.81
EmpDG	19.96	9.11	4.74	2.80	31.65	0.46	1.99	0.57	37.43
KEMP	18.07	8.30	4.37	2.65	36.40	0.55	2.29	0.52	36.89
DCKS	21.73	10.62	6.24	4.09	–	2.19	9.61	-	16.08
E-CORE	19.77	5.65	3.28	2.11	–	0.68	3.38	-	33.04

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.

Models	B-1 $\uparrow$	$\textbf{D-1}\uparrow$	D-2 $\uparrow$	PPL $\downarrow$
Transformer	-	1.29	6.91	81.55
MoEL	19.04	2.33	15.26	62.93
GLHG	19.66	3.50	21.61	15.67
Emstremo	20.96	2.90	14.80	-
EESC	21.38	4.88	25.95	14.88
Ours	22.03	5.11	29.63	10.09

Table 2: Automatic evaluation results on the ESConv.

Comparison	Aspects	Win	Lose	Tie
Ours vs. MOEL	Emp.	46.1	23.6	30.3
	Rel.	39.8	22.4	37.8
	Flu.	33.5	15.6	50.9
Ours vs. EmpDG	Emp.	51.6	18.5	29.9
	Rel.	50.3	15.3	34.4
	Flu.	35.1	13.5	51.4

Table 3: Human A/B evaluation results on the EMPATHETICDIALOGUES dataset.

Models	B-1 $\uparrow$	$\textbf{Acc}_{emo}\uparrow$	D-2 $\uparrow$	PPL $\downarrow$
Ours	24.66	57.57	13.68	9.26
w/o $e$	23.58	-	13.66	10.47
w/o $sum$	22.91	54.14	13.42	7.78
w/o $stra$	22.44	54.58	15.09	6.98
w/o CoT	22.55	-	13.26	8.64
w/o $R.\&S.$	22.22	55.93	15.25	11.86

Table 4: Evaluation results of the ablation study on the EMPATHETICDIALOGUES dataset.

R.\&S.

refer to the two modules: Consistency-Based Scoring and Ranking, and Strategy-Aware Sampling Scheme.

Models	B-1 $\uparrow$	$\textbf{Acc}_{emo}\uparrow$	D-2 $\uparrow$	PPL $\downarrow$
Ours	24.66	57.57	13.68	9.26
1/2 Data	23.43	56.35	15.07	12.75
1/4 Data	23.37	55.12	14.56	8.80
1/8 Data	23.28	51.29	14.50	9.70
1/16 Data	3.32	45.53	3.38	6.21
Zero Data	14.86	–	21.49	1.09

Table 5: Evaluation results for training set size and backbone model analyses. The abbreviation "Qwen3-4B-In." denotes Qwen3-4B-Instruct.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Strategy		Definition	Difficulty
1	Gratitude Prompting	Encourages the speaker to notice and reflect on positive experiences or supportive aspects of their life, fostering positive emotional awareness.	I
2	Restatement or Paraphrasing	Rephrases the speaker’s main ideas to demonstrate understanding and attentive listening.	I
3	Others	Covers responses that do not clearly fit into any predefined strategy category.	I
4	Information	Provides objective facts or relevant knowledge to help the speaker better understand their situation or make decisions.	I
5	Neutral Validation	Affirms that neutral or low-intensity emotional states are normal and acceptable without encouraging stronger emotions.	II
6	Positive Reinforcement	Highlights the speaker’s strengths, efforts, or constructive behaviors to reinforce confidence and motivation.	II

Appendix

Table 6: The Structured Empathy Strategy System of STRIDE-ED.

Context	Speaker : I am going to a concert very soon.
Summary	The speaker is anticipating an upcoming concert they will attend.
Emotion	Excited
Strategy	Exploring Actions and Intentions
Actions	I chose Exploring Actions and Intentions because the speaker’s excitement about the concert hints at unfolding plans, and asking about details channels that energy into shared anticipation.
Generated Response	That’s awesome! Who are you going to see?
Ground Truth	Cool! Who are you going to see?

Appendix

Table 7: Case Study 1 illustrating the use of the Cognitive Reframing strategy

Context	Speaker : 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 can’t believe it.
	Listener : That’s great!! History 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 he’s ever made. I honestly thought I had the next Einstein on my hands. Now I guess I’ll have to settle for…above average, I guess. :(
Summary	The speaker’s son achieved a B in history, his first grade below an A, which initially made the speaker hopeful but now feels disappointed.
Emotion	surprised
Strategy	Cognitive Reframing
Actions	I chose Cognitive Reframing because the speaker’s disappointment over a single B clashed with sky-high expectations, and recasting the grade as proof of challenging coursework softens the sting and restores perspective.

Appendix

Table 8: Case Study 2 illustrating the use of the Cognitive Reframing strategy

Context	Speaker : 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, especially the milestone years. Now my grandchildren are starting school. Time flies by too quickly
Summary	The speaker reflects on their children’s school years, now experiencing the same bittersweet emotions as their grandchildren begin school, expressing how quickly time passes.
Emotion	Sentimental
Strategy	Self-disclosure
Actions	I chose Self-disclosure because the speaker’s nostalgia over grandchildren starting school echoes my own disbelief at how fast kids grow, and sharing that sense of fleeting time affirms we’re in the same bittersweet boat.

Appendix

Table 9: Case Study 3 illustrating the use of the Self-disclosure strategy

Context	Speaker : I am going to my local amusement park tomorrow and feeling a certain way about it!
Summary	The speaker is anticipating a visit to an amusement park and feels a mix of excitement and anxiety about the experience.
Emotion	Apprehensive
Strategy	Exploring Thoughts and Cognition
Actions	I chose Exploring Thoughts and Cognition because the speaker’s mixed excitement-and-apprehension about the amusement park suggests hidden worries, and probing the thoughts behind the nerves can turn anxiety into eager curiosity.
Generated Response	That sounds like a lot of fun! Are you nervous?
Ground Truth	Which way are you feeling? Are you a thrill seeker?

Appendix

Table 10: Case Study 4 illustrating the use of the Exploring Thoughts and Cognition strategy

Context	Speaker : I’m so happy summer classes are officially over and I’ll be able to relax for a week!
	Listener : Great what will you be doing during your time off ?
	Speaker : I’m 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. It’s that school mentality.
Summary	The speaker is conflicted about balancing relaxation and productivity during their week-long break, struggling with the lingering mindset of being productive despite wanting downtime.
Emotion	Joyful
Strategy	Providing Suggestions
Actions	I chose Providing Suggestions because the speaker felt torn between rest and productivity, and offering concrete ways to enjoy downtime eases the guilt while keeping the momentum alive.

Appendix

Table 11: Case Study 5 illustrating the use of the Providing Suggestions strategy.

Context	Speaker : The litter of puppies had been born!
	Listener : Nice, amazing to witness birth.
	Speaker : All of the puppies survived, truly a lucky amazing occasion!
Summary	The speaker is celebrating the successful birth of their puppies and expressing gratitude for their survival.
Emotion	joyful
Strategy	Affirmation and Reassurance
Actions	I chose Affirmation and Reassurance because the speaker is deeply joyful about their puppies’ birth and expressing thanks for their survival; this reinforces their positive feelings and validates their relief by acknowledging the success and wishing them well.

Appendix

Table 12: Case Study 6 illustrating the use of the Affirmation and Reassurance strategy

Emotion-Aware Dialogue Generation

Emotion-aware dialogue generation adapts conversational responses to users’ feelings, context, and interaction goals across text, speech, and multimodal settings.

See all 54 papers

EchoChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue
Oct 4, 2026Dingdong Wang, Shujie Liu, Yayue Deng +6Emotion-Aware Dialogue Generation Speech Language Models
The Chinese University of Hong Kong, China · Microsoft Corporation
ER-EDF: A Psychology-Grounded Emotion Regulation Framework for Speech Empathetic Dialogue Generation in Large Audio-Language Models
Sep 14, 2026Hongyu Jin, Wenda Zhang, Runqiu Fei +3Spoken Dialogue Systems Audio-Language Models
PRIDE: Privileged Information-enhanced Distillation for Empathetic Dialogue Generation
Jun 22, 2026Jiaqiang Wu, Zhouan Zhu, Shangfei WangLanguage Model Distillation Learning Using Privileged Information
Anhui Robot Technology Standard Innovation Base, School of Computer Science and Technology, University of Science and Technology of China, Hefei, Anhui 230027, PR China

Oct 4, 2026J/K move · Enter open · S save

EchoChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue

Dingdong Wang, Shujie Liu, Yayue Deng, Yuxuan Hu+5

The Chinese University of Hong Kong, China · Microsoft Corporation

Emotion-Aware Dialogue Generation Speech Language Models RL for Language Model Reasoning

Open paper AskPDF arXiv

Empathetic spoken dialogue is a sophisticated cognitive process that requires not only recognizing emotions but also inferring a user's latent mental states to provide appropriate support. However, current SpeechLLMs often treat empathy as a direct input-to-response mapping, leading to "superficially warm" but emotionally hollow interactions. In addition, since empathy relies on a multi-stage process with strong inter-step dependency, errors at any intermediate step can cascade through subsequent steps and lead to inappropriate responses, while existing training paradigms lack mechanisms to precisely localize and improve such errors. In this work, we propose EchoChat, a unified framework that reformulates empathetic spoken dialogue as a structured cognitive reasoning process integrating perception, mental-state reasoning, and response generation. To support this paradigm, we first construct EchoDialogue-400K, an acoustically rich dataset for multi-stage empathetic supervision. During the SFT stage, we strengthen acoustic grounding through proposed Acoustic-Anchored Attention (AAA). During the RL stage, we further introduce a novel stage-aware optimization objective with Step-Decomposed Credit Assignment (SDCA) to localize reasoning errors and mitigate cascaded error propagation. In addition, we introduce EchoEval, an expert-annotated benchmark for multi-dimensional empathy evaluation. Extensive experiments demonstrate that EchoChat achieves state-of-the-art performance in perception, reasoning, and response alignment. Project page: https://github.com/dingdongwang/EchoChat

Abstract

Emotion-Aware Dialogue Generation

Explore similar work

EchoChat: Structured Cognitive Reasoning in Empathetic Spoken Dialogue

ER-EDF: A Psychology-Grounded Emotion Regulation Framework for Speech Empathetic Dialogue Generation in Large Audio-Language Models

PRIDE: Privileged Information-enhanced Distillation for Empathetic Dialogue Generation