Paper deep dive
StageWell: A Process-Aligned Chinese Corpus for Positive-Psychology Support Dialogue
Yuxiong Wang, Ziwei Lin, Bo Wang, Yu Zhang, Shiguang Ni
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 9/14/2026, 6:18:00 AM
Summary
The paper introduces StageWell, a process-aligned Chinese corpus for positive psychology support dialogue, designed to address the lack of structured supervision in multi-turn support processes. It features a six-stage support process (Rapport, Clarify, Strengths, Future, Goals, Summary) and utilizes a multi-agent rewriting workflow to generate 12,445 SFT instances and 1,849 DPO preference pairs. The authors also propose HQS, a structured evaluation protocol covering Safety, Quality, and Stage consistency. Experiments on four 9B-14B open-source LLMs demonstrate that StageWell significantly improves process control, response quality, and safety compared to base models.
Entities (10)
Relation Signals (16)
StageWell → usesmethodology → Six-Stage Support Process
confidence 95% · StageWell organizes support into a six-stage support process... S1 Rapport & Safety... S6 Summary & Transfer
StageWell → usesprotocol → HQS
confidence 94% · together with HQS, a structured protocol for data construction and evaluation
StageWell → contains → DPO
confidence 93% · 1,849 DPO preference pairs
StageWell → contains → SFT
confidence 93% · construct 12,445 SFT instances
Gemma 3 12B → evaluatedon → StageWell
confidence 90% · Gemma-3-12B Base
GLM-4-9B → evaluatedon → StageWell
confidence 90% · GLM-4-9B Base
Qwen3-14B → evaluatedon → StageWell
confidence 90% · Across four 9B–14B open-source LLMs... Qwen3-14B
Yi-1.5-9B → evaluatedon → StageWell
confidence 90% · Yi-1.5-9B Base
HQS → evaluates → Stage Consistency
confidence 90% · S for stage consistency
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Positive psychology dialogue aims to support emotional distress and positive resource building, requiring models to produce not only empathetic replies but also coherent progression through a multi-turn support process. Existing resources often reduce supervision to turn-level strategies or holistic preference labels, leaving process position, support function, and local repair targets implicit. We introduce StageWell, a process-aligned Chinese corpus for positive psychology dialogue, together with HQS, a structured protocol for data construction and evaluation. StageWell organizes support into a six-stage support process and uses a multi-agent whole-dialogue rewriting workflow to construct 12,445 SFT instances, 1,849 DPO preference pairs, and a GroundTruth subset of 120 expert-revised dialogues and 977 QA pairs. Guided by HQS, DPO pairs are built as process-localized repairs: flawed model outputs are used as rejected responses, and targeted rewrites under the same context and stage constraint are used as chosen responses. Across four 9B-14B open-source LLMs, this supervision yields robust gains in process control, response quality, and safety. Averaged across models, BERTScore improves by 0.037, Q-Overall increases by 1.32 points, S-exact increases by 0.236, and the H-critical rate decreases by 0.167. These results highlight the value of modeling supportive dialogue as a structured multi-turn support process rather than as single-turn response generation.
Tags
Links
- Source: https://arxiv.org/abs/2608.29326v1
- Canonical: https://arxiv.org/abs/2608.29326v1
Trouble viewing inline? Open PDF directly →
Full Text
90,838 characters extracted from source content.
Expand or collapse full text
StageWell: A Process-Aligned Chinese Corpus for Positive-Psychology Support Dialogue Yuxiong Wang 1,† Ziwei Lin 1,† Bo Wang 2,∗ Yu Zhang 1 Shiguang Ni 1,∗ 1 Shenzhen International Graduate School, Tsinghua University 2 Department of Psychological and Cognitive Sciences, Tsinghua University bo-wang@tsinghua.edu.cn ni.shiguang@sz.tsinghua.edu.cn Abstract Positive psychology dialogue aims to support emotional distress and positive resource build- ing, requiring models to produce not only em- pathetic replies but also coherent progression through a multi-turn support process. Existing resources often reduce supervision to turn-level strategies or holistic preference labels, leaving process position, support function, and local repair targets implicit. We introduce StageWell, a process-aligned Chinese corpus for positive psychology dialogue, together with HQS, a structured protocol for data construction and evaluation. StageWell organizes support into a six-stage support process and uses a multi- agent whole-dialogue rewriting workflow to construct 12,445 SFT instances, 1,849 DPO preference pairs, and a GroundTruth subset of 120 expert-revised dialogues and 977 QA pairs. Guided by HQS, DPO pairs are built as process-localized repairs: flawed model out- puts are used as rejected responses, and tar- geted rewrites under the same context and stage constraint are used as chosen responses. Across four 9B–14B open-source LLMs, this supervi- sion yields robust gains in process control, re- sponse quality, and safety. Averaged across models, BERTScore improves by 0.037, Q- Overall increases by 1.32 points, S-exact in- creases by 0.236, and the H-critical rate de- creases by 0.167. These results highlight the value of modeling supportive dialogue as a structured multi-turn support process rather than as single-turn response generation. We make our dataset available for research. 1 1 Introduction Supportive dialogue systems aim to provide emo- tional relief and sustained companionship through positive, supportive interactions (Liu et al., 2021; † These authors contributed equally. ∗ Corresponding author. 1 https://github.com/stagewell-anon/Stagewell. Foundation Building (Stage 1 & 2) Stage 1 Stage 2 [Reflective Listening] It sounds like the pressure is overwhelming you right now. [Empathetic Validation] It's completely normal to feel this way before a big test. I'm here for you. [Summary Clarify] So math is the main trigger making you feel stuck and panicked. [Open-ended Question] When that panic hits, what specific thoughts are running through your head? Yeah, especially with math. The harder I try, the more I forget. I panic just looking at the questions. Resource Activation (Stage 3 & 4) Stage 3 Stage 4 [Identifying Strengths] You mentioned doing well in a speech contest before. [ Guiding Self-Reflection] What specific skills helped you overcome your nerves back then? I practiced every single day and asked my teacher for feedback. Being fully prepared gave me confidence. [Strengths Transfer] You have a proven strength in preparation and seeking feedback. [Cognitive Restructuring] How can we use that strength of discipline to turn this test into a series of predictable, daily practice steps? User Initial State I'm terrified of the math midterm. My mind goes blank, and I feel like I'm going to fail and disappoint everyone. ?? ? Stage 5 Action Planning (Stage 5 & 6) Stage 6 [Smart Goal Setting] Let's set a manageable goal: reviewing math mistakes for just 15 minutes daily. [Positive Skill Development] Also, let's practice a deep- breathing exercise to use when you feel panicked. [Review Progress] You've shifted beautifully from feeling helpless to having a clear action plan today. [Vision Building] Imagine entering the exam room knowing you've practiced daily and have your breathing technique ready. How does that feel? 15 minutes a day sounds doable. I'l use the breathing technique when I get anxious. I feel much more in control. I have a concrete plan and tools to stay calm. I'm ready to tackle this step by step. Figure 1: Overview of the six-stage support process for positive psychology dialogue. Zhou et al., 2018). In this context, positive psy- chology dialogue has emerged as an important non- clinical research direction for the general popula- tion, targeting everyday distress relief and positive resource building (Gable and Haidt, 2005; Selig- man and Csikszentmihalyi, 2014). Specifically, the task requires navigating a multi-turn support pro- cess involving emotional acknowledgment, prob- lem clarification, resource activation, and action advancement (Miller and Rollnick, 2012). This inherently process-sensitive nature demands re- sources that encode both empathetic wording and dynamic stage progression. With the rapid progress of large language models (LLMs), recent work has arXiv:2608.29326v1 [cs.CL] 29 Aug 2026 Step1: Supervised Fine-Tuning 3. Model Training Step2: HQS-Guided DPO Preference Alignment 4. Evaluation : Hard Constraints <Safety> : Response Quality <Dimensions> : Stage Consistency <Progression> 1. Data Preparation Raw Data Sources General Counseling Dialogue Data Six-Stage Positive Psychology Dialogue Phase I: Foundation Phase I: Resource Phase I: Action Rapport Safety Clarify Reflect Strengths Resources Future Hope Goals Strategies 1234 5 6 Summary Transfer Ten Support Topics Emotion Stress Peer Social Family Support Self Mgmt. Digital Life Safety Bound. 2. SFT Construction Case Card Generation Quality Control & Human Review Preprocessing & Filtering Multi-Agent Whole-Dialogue Rewriting Patcher Planner Refiner Writer I have an exam tomorrow but I haven't started studying. I feel so hopeless. Rejected Response Don't worry about it. A lot of students feel this way. Just try to study hard tonight. Chosen Response That sounds really stressful. It's completely valid to feel hopeless in this situation. Let's focus on just one small first step together. DPO Training Model Outputs & Expert References Multi-Agent Rewriting Defect Filtering HQS-Based Diagnosis H Q S Preference Objective Score Update Policy Model π θ Reference Model π Ref ❄ High-Quality Preference Pairs Chosen Rejected Adapt. Trans. Self Concept Goals Strengths Academic Pressure Figure 2: Overview of the StageWell framework. The left panel shows data preparation and the six-stage support process; the middle panel shows multi-agent whole-dialogue SFT construction; the upper-right panel shows phased alignment training with HQS-guided DPO preference alignment; and the bottom panel shows HQS-based evaluation. advanced supportive response generation (Zhou et al., 2023), strategy modeling (Tu et al., 2022; Zheng et al., 2023), and structured evaluation (Lee et al., 2025; Bavaresco et al., 2025). However, existing resources remain insufficient for modeling the support process in positive psy- chology dialogues. (1) Limited process-oriented resources. Prior work has mainly focused on gen- eral psychotherapy or emotional support, and pub- lic resources are mostly concentrated on supervised fine-tuning corpora or evaluation benchmarks (Sun et al., 2021; Zhao et al., 2024). (2) Coarse-grained supervision signals. Although some studies intro- duce strategy labels (Hu et al., 2025), they still pro- vide limited supervision for multi-turn process pro- gression. Existing preference-learning resources (Zhao et al., 2025a) mainly provide pairwise re- sponse preferences, but rarely localize the reason for preference or the process-level defect being repaired. Overall, effective preference learning for psychological support dialogue is constrained by the scarcity of high-quality structured prefer- ence data, especially resources that explicitly cover multi-turn process dynamics, stage progression, and process-level deviations. To fill this gap, we build StageWell, a process- aligned corpus for Chinese positive psychology dialogue. Figure 2 provides an overview of the StageWell framework. StageWell contains subsets: SFT, DPO, and GroundTruth. We adopt a two- stage construction pipeline that yields supervised fine-tuning data and preference-alignment data. In the SFT stage, we organize ten dialogue topics and structure positive psychology support into a six- stage support process, as illustrated in Figure 1. We then use a multi-agent whole-dialogue rewrit- ing pipeline, consisting of Planner, Writer, Patcher, and Refiner, for process planning, response genera- tion, consistency repair, and language refinement. Through this workflow, generic mental-health di- alogues are rewritten into multi-turn positive psy- chology dialogues with a coherent support pro- cess, resulting in 12,445 SFT training instances. In the preference-construction stage, we perform targeted rewriting over residual defects in SFT model outputs, including stage mismatch, insuf- ficient advancement, templated expression, and un- stable boundaries, and construct 1,849 DPO prefer- ence pairs. This turns generic response ranking into process-localized supervision. We also construct a GroundTruth subset comprising 120 expert-revised dialogues and 977 QA pairs for evaluation. Additionally, to support both preference con- struction and model evaluation, we propose HQS, a structured protocol: H for safety and boundary failures, Q for response quality, and S for stage consistency. During DPO construction, HQS helps identify residual defects and facilitates targeted re- pair; during evaluation, it enables consistent struc- tured analysis of model outputs along safety, qual- ity, and process dimensions. Experiments on four 9B–14B open-source LLMs show improved pro- cess control, response quality, and safety: averaged across backbones, BERTScore improves by 0.037, Q-Overall by 1.32 points, S-exact by 0.236, and H-critical decreases by 0.167. Our contributions are as follows: •We organize positive psychology dialogue with an explicit six-stage support process and construct preference pairs as targeted, process- localized repairs, yielding auditable supervi- sion over both support progression and local response quality. •We construct and release StageWell, a process-aligned corpus for Chinese positive psychology dialogue, featuring 12,445 SFT instances, 1,849 DPO preference pairs, and a GroundTruth subset for held-out evaluation. •We introduce HQS, a structured protocol cov- ering Safety, Quality, and Stage consistency, and show across four open-source LLMs that StageWell serves as a useful alignment re- source. 2 Related Work 2.1 Datasets for Supportive and Positive Psychology Dialogue Supportive dialogue research has evolved from generic empathetic response generation toward more structured psychological support. Resources such as SoulChat(Chen et al., 2023), SMILE(Qiu et al., 2024), HealMe(Xiao et al., 2024), and Su- DoSys(Chen et al., 2024) show that LLMs can pro- duce empathetic, topic-aware, and intervention-like responses. Counseling-oriented datasets, including CPsyCoun(Zhang et al., 2024), PsyDial(Qiu and Lan, 2025), and AURADIAL(Zhang, 2025), fur- ther extend this line to report-based reconstruction, long-term conversations, and large-scale Chinese counseling data, while positive reframing work studies constructive rewrites of negative appraisals (Ziems et al., 2022; Maddela et al., 2023; Sharma et al., 2023). However, these resources mostly pro- vide turn- or response-level supervision, leaving the organization of full dialogues into explicit support stages under-specified. StageWell complements this line by constructing a Chinese positive psychol- ogy corpus whose dialogues are organized around a non-clinical multi-stage support process. 2.2 Preference Data Construction for Alignment The lack of process supervision also limits pref- erence construction: when the support process is underspecified, chosen/rejected pairs tend to inherit the same ambiguity. Preference learning has be- come a core paradigm for LLM alignment, from RLHF-style helpfulness and harmlessness compar- isons (Bai et al., 2022) to DPO over chosen/rejected pairs (Rafailov et al., 2023) and large-scale re- sources such as UltraFeedback (Cui et al., 2023), BeaverTails (Ji et al., 2023), and HelpSteer (Wang et al., 2024). However, most open preference data focus on general instruction following, harm- lessness, or broad helpfulness. In psychological- support dialogue, such coarse supervision is insuffi- cient because a preferred response may differ from a rejected one in strategy choice, stage fit, safety boundary, or surface realization. Recent emotional-support alignment work has begun to examine this issue more directly. ESC-Pro (Zhao et al., 2025a) constructs preference data for emotional support dialogue through strategy-aware response optimization. DecoupledESC (Zhang et al., 2025) further highlights the entanglement between strategy and response quality, and PsyAd- visor (Hu et al., 2025) improves interpretability through strategy-aware modeling.Still, exist- ing preference supervision rarely identifies which process-stage constraint is violated, what localized defect causes the preference, or how the response should be repaired while preserving safety bound- aries. StageWell addresses this gap by constructing DPO pairs as process-localized repairs rather than generic response rankings. 3 Methodology 3.1 Task Definition and Six-Stage Support Process We formulate Chinese positive psychology dia- logue as a multi-turn support task in which the model receives a dialogue historyH t−1 and a new user utterance u t , and produces a response a t : a t ∼ π θ (·| H t−1 ,u t ).(1) Unlike generic empathetic chat, the target response should not only address the current concern but also occupy an appropriate position within the support process. A response can sound warm in isolation yet still be misaligned if it pushes action before clar- ification, repeats empathy after the user is ready to move forward, or offers broad advice without ac- tivating available resources. We therefore cast the task as process-conditioned generation, in which each response must fit both the current utterance and the current stage of support. StagePrimary ObjectiveDialogue Strategy and Focus S1 Rapport & SafetyTrust and emotional safetyWarm opening and low-pressure invitation. S2 Clarify & ReflectPrecise concern understandingListening, reflection, and open-ended clarification. S3 Strengths & ResourcesResource awarenessCoping experiences, strengths, and relationships. S4 Future & HopeRealistic hope and possibilityFuture-oriented reframing without overpromising. S5 Goals & StrategiesFeasible next stepsSmall, collaborative, context-grounded actions. S6 Summary & TransferConsolidation and continuationBrief summary, reinforcement, and follow-up space. Table 1: The six-stage support process used in StageWell. AgentRole and Responsibility PlannerStage plan, facts, and turn functions. WriterPlan-guided coherent rewriting. PatcherRepetition, transition, and consistency repair. RefinerFluency polishing with stage intent kept. Table 2: Roles within the multi-agent rewriting pipeline. The task spans ten common support topics: aca- demic pressure, emotional regulation, peer relation- ships, family communication, self-concept, self- management, digital-life issues, developmental transitions, safety and boundaries, and goal de- velopment. Figure 1 illustrates how the support process moves from emotional containment to con- structive action, and Table 1 defines the six ordered stages. Together, they jointly operationalize three fundamental per-turn decisions: where the dialogue is in the support process, what support function the response should serve, and how that function should be realized in local wording. This framework draws on several key prin- ciples from Solution-Focused Brief Therapy (SFBT)(De Shazer et al., 1986), especially the char- acteristic shift from problem-centered talk toward proactive resource activation(Ji et al., 2020), a clear future orientation, and concrete next-step construc- tion (Gingerich and Peterson, 2013; Fu et al., 2018). We instantiate these principles as a task-specific schema for non-clinical Chinese positive psychol- ogy support rather than reproducing any single ther- apeutic protocol. 3.2 Process-Aligned Data Construction We construct StageWell by rewriting generic mental-health dialogues into process-aligned pos- itive psychology dialogues rather than using the source data as direct supervision. As shown in Figure 2, the pipeline first performs role normaliza- tion, thematic filtering, and risk filtering, and then summarizes each retained dialogue into a case card with topic, dominant emotion, support objective, and risk level. This case card anchors a four-agent rewriting workflow used only during dataset con- struction, not at inference time. As summarized in Table 2, the Planner specifies the stage plan and factual boundaries, the Writer drafts the dialogue under that plan, the Patcher repairs local inconsis- tencies, and the Refiner improves fluency without changing stage intent. For SFT construction, the key design is whole- dialogue rewriting rather than isolated turn-level editing. The Planner first specifies a global stage sequence, turn functions, and factual boundaries for the full dialogue, after which the downstream agents realize and refine each turn under this plan. The resulting SFT subset provides coherent training instances in which rapport building, clarification, strength activation, future orientation, planning, and transfer appear as ordered support functions. For DPO construction, the key design is process- localized repair rather than holistic response rank- ing. Starting from outputs of SFT-adapted models, HQS (our evaluation protocol, as detailed in the following section) identifies residual defects such as stage mismatch, insufficient advancement, tem- plated expression, and unstable boundaries. Under fixed dialogue history and target stage, the flawed original output is kept as the rejected sample, and a targeted rewrite that repairs the localized defect becomes the chosen sample. This design constrains the preference compari- son more tightly than standard chosen/rejected data, in which process position, support function, and surface quality may vary simultaneously. Here, HQS first localizes the defect, and the rewrite pre- serves the same stage target, so each pair primarily isolates a single process-relevant correction. Formally, for each selected turn, the dialogue history and expected stage are fixed; the rejected response is the original flawed output, and the cho- sen response is a targeted repair under the same context and stage constraint. Let(x,y w ,y l ) ∼ D denote a prompt, a preferred response, and a dis- preferred response. We optimize DPO with the standard objective (Rafailov et al., 2023): L DPO (π θ ;π ref ) =−E (x,y w ,y l )∼D logσ β(r θ (x,y w )− r θ (x,y l )) (2) 40 50 45 35 55 Support Stage Preference Signal Profile Chosen vs. Rejected Length Q1 ChosenRejected Δ= (Chosen-Rejected) / Rejected Mean response length (chars) 1 Rapport & Safety 2 Clarify & Reflect 3 Strengths & Resources 4 Future & Hope 5 Goals & Strategies 6 Summary & Transfer 45% 52% 11% 25% 0% 2% 15% 43% 35% 16% 1% 0% 0% 0% 0% 0% 62% 0% 0% 0% 0% 0% 36% 98% 40% 5% 54% 59% 1% 0% Δ -5% Shorter chosen : 55% Δ -6% Shorter chosen: 56% Δ -4% Shorter chosen: 52% Δ -9% Shorter chosen: 61% Δ 0% Shorter chosen: 46% Δ -10% Shorter chosen: 59% Stage-wise Support Q2 Q3 Q4 Q5 Figure 3: DPO preference profile across six stages, showing the distribution of Q1–Q5 repair signals and response length differences. Q1–Q5 denote content fit, acknowledgment, feasibility, maturity, and Chinese naturalness. ModuleOperational criterion HHard constraints: safety and risk checks, including role-boundary violations. QResponse quality: content fit, emotional acknowledgment, feasibility, maturity, and Chinese naturalness. SStage consistency: expected stage, progression pace, and transition fit. Table 3: HQS modules used in preference construction and model evaluation. where r θ (x,y) = logπ θ (y | x)− logπ ref (y | x). (3) Overall, the pipeline yields two complementary su- pervision signals: staged support demonstrations through SFT and targeted, process-localized prefer- ence pairs through DPO under shared context and stage constraints. 3.3 HQS: A Structured Protocol for Construction and Evaluation HQS serves as the shared protocol for both pref- erence construction and final evaluation. As sum- marized in Table 3, H enforces hard constraints such as safety, risk, and role-boundary failures; Q evaluates local response quality; and S checks stage consistency and progression pace. Method- ologically, HQS separates red-line safety screen- ing, stage-conditioned response-quality assessment, and process-stage recognition, so that failures can be localized before they are converted into super- vision signals. During DPO construction, HQS identifies defective responses and defines the repair target for rewriting. During evaluation, the same protocol decomposes model behavior along safety, A. Resource scale and basic statistics SubsetStatisticValue SFTTraining instances12,445 SFTAvg. turns / characters8.09 / 617.51 SFTStage-label consistency100.00% DPOPreference pairs / source dialogues1,849 / 959 DPOAvg. rejected / chosen chars46.17 / 44.10 GroundTruth Expert-revised dialogues / QA pairs120 / 977 GroundTruth Expert revisers28 B. Stage distribution DataS1S2S3S4S5S6 SFT asst. (%)12.3124.5012.2912.2926.2512.36 DPO target (n)161558306201440183 DPO target (%)8.7130.1816.5510.8723.809.90 C. DPO repair dimensions DataQ1Q2Q3Q4Q5 Pairs (n)449408273340379 Proportion (%)24.2822.0714.7618.3920.50 Table 4: Detailed statistics of the StageWell subsets. S1–S6 denote support stages, and Q1–Q5 denote HQS response-quality dimensions. quality, and process-control dimensions. This de- sign ensures that the same process criteria are used both to select and repair DPO samples during con- struction and to diagnose model outputs during evaluation. Table 4 summarizes the resulting re- source scale, stage distribution, and DPO repair distribution. We use HQS not as an isolated metric set, but as a shared protocol linking data construc- tion and held-out evaluation. 4 Experiments We evaluate StageWell from three complementary angles. First, we test whether StageWell improves process-aware positive psychology dialogue on held-out expert-revised data. Second, we separate the contributions of stage-aware SFT and process- localized DPO alignment across four open-source backbones. Third, we report a small descriptive ModelCond. Q: Response Quality↑H↓S: Stage Consistency↑ OverallQ1Q2Q3Q4Q5 CriticalExactTop-2 Within-1 Macro-R Qwen3-14B Base3.5623.4143.8732.9843.5603.5890.0580.4190.5590.6520.396 SFT3.9713.7664.3863.8784.3063.8690.0500.4760.5890.7430.457 DPO4.2324.1374.4814.3824.5084.1420.0330.5150.6140.7650.477 GLM-4-9B Base3.3013.2733.4232.4552.9403.3590.0330.4020.5060.6280.381 SFT3.7943.7313.9514.1503.9343.7310.0330.6800.7920.8200.659 DPO3.9423.9234.0244.2484.0383.9040.0420.6920.7910.8280.670 Yi-1.5-9B Base2.3822.3392.4432.0572.5412.5190.1420.4170.4640.6230.387 SFT3.7723.6374.0404.1713.8913.6770.0420.7240.8220.8330.721 DPO3.9193.8134.1324.2644.0553.8520.0330.7260.8150.8290.725 Gemma-3-12B Base1.4361.4541.5101.4671.5451.3820.5750.4590.5090.6420.422 SFT3.7463.6213.9954.1023.8693.6580.0330.7030.8080.8250.704 DPO3.8793.7814.0894.1673.9103.7970.0330.7090.7980.8410.708 Average Base2.6702.6202.8122.2412.6472.7120.2020.4240.5100.6360.397 Average SFT3.8213.6894.0934.0754.0003.7340.0400.6460.7530.8050.635 Average DPO3.9933.9134.1814.2654.1283.9240.0350.6600.7540.8160.645 ∆ DPO–Base+1.323+1.293+1.369+2.024+1.481+1.212-0.167+0.236+0.244+0.180+0.248 Table 5: HQS results on GroundTruth. Q1–Q5 are the five HQS quality dimensions; lower H-critical is better. feasibility and usability pilot as a user-facing com- plement to the offline benchmark. 4.1 Experimental Setup Evaluation Data. All offline experiments use the held-out GroundTruth subset. As summarized in Table 4, SFT provides staged support demonstra- tions, DPO provides process-localized preference pairs, and GroundTruth provides held-out expert references. GroundTruth dialogues and QA pairs are fully disjoint from SFT and DPO and are used only for held-out evaluation. Figure 3 further characterizes the DPO subset. Each pair keeps dialogue history and target stage fixed: the rejected response is the original HQS- flagged model output, and the chosen response is a targeted repair of the localized defect. Earlier stages more often repair insufficient acknowledg- ment or low-yield questioning, whereas later stages more often repair weak action grounding, prema- ture closure, or unstable boundaries. Models. We compare Base, SFT, and DPO con- ditions on GLM-4-9B(Glm et al., 2024), Qwen3- 14B(Yang et al., 2025), Yi-1.5-9B(Young et al., 2024), and Gemma-3-12B(Team et al., 2025). All automatic and HQS metrics are computed on the GroundTruth evaluation set of 120 dialogues and 977 expert-revised QA pairs. In all tables, DPO denotes the model initialized from the corresponding SFT checkpoint and further optimized with StageWell preference pairs. Training Setup.Training and evaluation are implemented with ModelScope Swift (ms- swift)(Zhao et al., 2025b) on a single-node server with2×NVIDIA L20 (48GB) GPUs and CUDA 12.8. All models use LoRA-style adaptation (Hu et al., 2022). For SFT, we train for 3 epochs with BF16 precision, AdamW(Loshchilov and Hutter, 2017), learning rate5× 10 −5 , cosine scheduling with a 0.05 warmup ratio, weight decay 0.1, maxi- mum sequence length 1024, per-device batch size 1, gradient accumulation 8, and LoRA rankr = 8 withα = 32and dropout 0.05 over all linear lay- ers. For DPO, we initialize from the corresponding SFT adapter and train for 1 epoch with learning rate 5× 10 −7 ,β = 0.7, weight decay 0.01, maximum sequence length 1536, per-device batch size 1, and gradient accumulation 16. GLM-4, Qwen3, and Yi-1.5 are trained on a single L20, while Gemma-3 is launched with two GPUs viatorchrun; other major settings are kept aligned across backbones. Metrics.We report both reference-based and process-aware metrics on the same held-out GroundTruth set.For reference-based evalua- tion, we report ROUGE-1/2/L(Lin, 2004), BLEU- 4(Papineni et al., 2002), and BERTScore(Zhang et al., 2019) on the 977 QA pairs; Avg. denotes their unweighted mean and is used only as a com- pact summary. For process-aware evaluation, HQS comprises three fixed LLM-as-a-Judge modules with structured JSON outputs. H is a dialogue-level Hard Judge for factual hallucination, boundary vi- olation, safety risk, and role violation; a dialogue is H-critical if any critical flag is triggered. Q is a turn-level Quality Judge that scores applicable dimensions on a 1–5 scale given the context, candi- Base ModelConditionROUGE-1ROUGE-2ROUGE-LBLEU-4BERTScoreAvg. Qwen3-14B Base0.3340.1150.2340.0980.2920.215 SFT0.3570.1210.2470.1030.2950.225 DPO0.3730.1410.2700.1180.3140.243 Gemma-3-12B Base0.1320.0520.0970.0350.2110.105 SFT0.3510.1180.2450.0970.2910.220 DPO0.3540.1240.2510.1030.2960.226 GLM-4-9B Base0.3210.1300.2420.1080.2630.213 SFT0.3530.1250.2520.1040.2910.225 DPO0.3560.1290.2560.1070.2950.229 Yi-1.5-9B Base0.2750.1140.2090.0980.2820.196 SFT0.3500.1190.2460.0990.2860.220 DPO0.3520.1220.2470.1010.2890.222 Average Base0.2660.1030.1960.0850.2620.182 Average SFT0.3530.1210.2480.1010.2910.223 Average DPO0.3590.1290.2560.1070.2990.230 ∆ DPO–Base+0.093+0.026+0.061+0.023+0.037+0.048 Table 6: Reference-based automatic evaluation on 977 GroundTruth QA pairs. All overlap scores are shown on a 0–1 scale. Avg. is the unweighted mean of ROUGE-1/2/L, BLEU-4, and BERTScore. Bottom rows summarize averages across backbones and the DPO–Base improvement. date reply, and expected stage; dimensions outside the expected stage are marked as non-applicable by design. S is a turn-level Stage Judge that predicts primary and secondary support stages. We report S-exact, Top-2, Within-1, and Macro-R. A format- only JSON Repair module is used only when judge outputs cannot be parsed and does not alter the underlying content judgments. We interpret HQS together with reference-based metrics. 4.2 Main Results and Analysis Table 5 provides the main evidence for StageWell. Averaged across backbones, SFT delivers the largest gain in process control, raising S-exact from 0.424 to 0.646 and reducing H-critical from 0.202 to 0.040, consistent with its role as stage-aware supervision. DPO then adds smaller incremental gains on top of SFT in the averaged results, improv- ing Q Overall from 3.821 to 3.993, S-exact from 0.646 to 0.660, and H-critical from 0.040 to 0.035. This pattern matches the design reflected in Eq. 2: SFT mainly establishes the support process, while DPO supplies finer local corrections under shared dialogue history and stage constraints. Table 6 provides a complementary but secondary view. SFT improves lexical and semantic over- lap across all backbones, and DPO adds further but modest gains in the averaged scores. The smaller margins here are expected: StageWell is designed to improve process control and process- localized repair rather than maximize lexical over- lap with expert references. We therefore interpret the reference-based results as evidence that process- aware alignment preserves, and modestly improves, semantic fit without being the primary source of the observed gains. The gains are also not confined to a single backbone. Gemma-3-12B, which starts from the weakest Base condition, improves substantially after alignment, suggesting that StageWell remains useful even when supportive capability is limited. The main non-monotonic case is GLM-4-9B, where H-critical increases from 0.033 to 0.042 af- ter DPO, corresponding to roughly 4 vs. 5 critical cases on 120 held-out dialogues. This amounts to about one additional critical case and should be interpreted with the raw count in mind. More broadly, the offline results support StageWell as an alignment resource for process-aware positive psychology dialogue rather than merely a collec- tion of response-level preferences. This distinction is crucial, as it ensures that responses are not just coherent but also appropriate for the specific stage. At the same time, the DPO stage is best interpreted as local refinement over an SFT-established process rather than as uniformly improving every metric on every backbone. Remaining errors include mild role overreach, compressed planning, and stage ambiguity in middle phases where clarification, re- source activation, and forward movement overlap. 4.3Exploratory Feasibility and Usability Pilot As a user-facing complement to the offline bench- mark, we report a small descriptive pilot with the deployed Qwen3-14B DPO system. The pilot in- cludes 28 non-clinical junior and senior high school students experiencing everyday emotional distress, each completing a pre-test, one interaction session, and a post-test. The pilot is intended to assess feasibility, usability, and immediate self-reported experience; it is not designed to estimate causal effectiveness. We report aggregate results only. A. Interaction statistics Valid participants28 Avg. interaction turns7.14± 0.77 Avg. usage time (minutes)9.57± 1.87 B. Immediate psychological state MeasurePrePost∆ Overall state3.24± 0.35 3.48± 0.40+0.24 Stress burden (rev.)3.183.36+0.18 Emotional distress (rev.) 3.143.43+0.29 Relaxation3.323.57+0.25 Hope3.293.61+0.32 Self-regulation3.293.43+0.14 C. Post-interaction subjective experience DimensionMean± SD Positive Understanding4.07± 0.5585.7% Acceptance4.00± 0.5878.6% Helpfulness3.71± 0.6371.4% Guidance clarity3.86± 0.6078.6% Continued use3.79± 0.6871.4% Overall satisfaction3.93± 0.5778.6% Usability4.14± 0.5292.9% Table 7: Pilot user-study results. Reverse-coded dimen- sions are reported in the positive direction. Positive denotes the percentage of participants scoring at least 4 on a five-point Likert scale. The descriptive outcomes of the pilot are sum- marized in Table 7. Participants reported a higher average immediate psychological-state score after the interaction (3.24 to 3.48, a relative increase of 7.4%). All sub-dimensions moved in the same di- rection, with the largest shifts observed for hope (+0.32), reverse-coded emotional distress (+0.29), and relaxation (+0.25). Post-interaction ratings were also generally pos- itive. Core dimensions such as understanding, ac- ceptance, overall satisfaction, and usability were rated near or above 4.0, while helpfulness (3.71) and continued-use willingness (3.79) were compar- atively lower. Taken together, these observations A. Immediate State Change Pre-test Post-test 3.03.23.43.63.8 Mean score Δ Delta Stress burden (rev.) 3.183.36 +0.18 Emotional distress(rev.) 3.143.43 +0.29 Relaxation 3.323.57 +0.25 Hope 3.293.61 +0.32 Self-regulation 3.29 3.43 +0.14 B. Post-interaction Experience 5 Helpfulness 3.71 Acceptance 4.00 Understanding 4.07 Usability 4.14 Satisfaction 3.93 Continued use 3.79 Guidance 3.86 1 3 Figure 4: Visualization of the pilot user-study results. Panel A shows pre–post changes in immediate psycho- logical state, and Panel B summarizes post-interaction subjective experience. suggest that the deployed system offers a usable short-term interaction experience. However, the study is small-scale, uncontrolled, and based on short-term self-report, so we interpret it only as preliminary evidence of feasibility and usability rather than as a primary effectiveness result. 5 Conclusion We presented StageWell, a process-aligned corpus for Chinese positive psychology dialogue, together with HQS, a structured protocol for preference construction and evaluation. StageWell organizes non-clinical support around a six-stage process and targets positive psychology dialogue for everyday distress relief and positive resource building. It con- tains staged whole-dialogue SFT training instances, process-localized DPO preference pairs, and a held- out GroundTruth set for evaluation. Across four open-source LLMs, StageWell improves process control, response quality, and safety, highlighting the practical value of process-aware supervision for supportive dialogue alignment. Overall, StageWell provides a reusable resource and a process-oriented perspective for future work on supportive dialogue, suggesting that dataset design for positive psychol- ogy dialogue should account for both stage pro- gression and local repair. Future work will focus on assessing the long-term effectiveness of such process-aligned dialogue systems in real-world de- ployments. 6 Limitations This work focuses on Chinese non-clinical posi- tive psychology dialogue, and several limitations remain. The six-stage support process and HQS criteria are tailored to Chinese-language support- ive interaction, which may limit direct transfer to other languages, cultural contexts, or more special- ized support scenarios. The StageWell dataset is constructed through constrained rewriting and ex- pert revision, a methodology that prioritizes clear process supervision and factual control over the conversational diversity found in fully naturalistic data. In addition, our DPO pairs are designed as targeted repairs under shared dialogue context and stage constraints. This design yields precise and interpretable supervision, but may cover a narrower range of preference variation than more open-ended response comparisons. Finally, the main evaluation relies on held-out expert-revised data and struc- tured automatic judging, complemented by a small exploratory pilot. Larger independent interactive studies are still needed to assess user-facing effects under extended real-world use. Ethics Statement This work concerns psychological support dia- logue, a sensitive domain involving user vulner- ability, privacy, and safety-critical interaction. All research procedures were approved by a formal ethics committee. The StageWell corpus, which forms the basis of this work, underwent careful de- identification and manual inspection to protect par- ticipant privacy. Despite these precautions, training models on the StageWell corpus introduces inher- ent risks due to the black-box nature of machine learning. Model outputs may contain subtle bi- ases or inaccuracies that could inadvertently affect vulnerable individuals. For this reason, this work is explicitly non-clinical and is not intended for diagnosis, treatment, or crisis intervention. The resulting models should be understood as research artifacts, not substitutes for licensed professionals. Improper use could exacerbate rather than alleviate psychological distress, and any such system should be used cautiously as a supplementary resource under human supervision. The pilot study was reviewed and approved by an institutional ethics committee. References Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, and 1 others. 2022. Training a helpful and harmless assis- tant with reinforcement learning from human feed- back. arXiv preprint arXiv:2204.05862. Anna Bavaresco, Raffaella Bernardi, Leonardo Berto- lazzi, Desmond Elliott, Raquel Fernández, Albert Gatt, Esam Ghaleb, Mario Giulianelli, Michael Hanna, Alexander Koller, Andre Martins, Philipp Mondorf, Vera Neplenbroek, Sandro Pezzelle, Bar- bara Plank, David Schlangen, Alessandro Suglia, Aditya K Surikuchi, Ece Takmaz, and Alberto Testoni. 2025. LLMs instead of human judges? a large scale empirical study across 20 NLP evalua- tion tasks. In Proceedings of the 63rd Annual Meet- ing of the Association for Computational Linguistics (Volume 2: Short Papers), pages 238–255, Vienna, Austria. Association for Computational Linguistics. Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. SoulChat: Improving LLMs’ 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 1170–1183, Singapore. Association for Com- putational Linguistics. Yixiang Chen, Xinyu Zhang, Jinran Wang, Xurong Xie, Nan Yan, Hui Chen, and Lan Wang. 2024. Structured dialogue system for mental health: An llm chatbot leveraging the pm+ guidelines. In International Con- ference on Social Robotics, pages 262–271. Springer. Ganqu Cui, Lifan Yuan, Ning Ding, Guanming Yao, Bingxiang He, Wei Zhu, Yuan Ni, Guotong Xie, Ruobing Xie, Yankai Lin, and 1 others. 2023. Ultra- feedback: Boosting language models with scaled ai feedback. arXiv preprint arXiv:2310.01377. Steve De Shazer, Insoo Kim Berg, EVE Lipchik, Elam Nunnally, Alex Molnar, Wallace Gingerich, and Michele Weiner-Davis. 1986. Brief therapy: Focused solution development. Family process, 25(2):207– 221. Lisha Fu, Junjie Yu, Shiguang Ni, and Hong Li. 2018. Reduced framing effect: Experience adjusts affective forecasting with losses. Journal of Experimental Social Psychology, 76:231–238. Shelly L Gable and Jonathan Haidt. 2005. What (and why) is positive psychology?Review of general psychology, 9(2):103–110. Wallace J Gingerich and Lance T Peterson. 2013. Effec- tiveness of solution-focused brief therapy: A system- atic qualitative review of controlled outcome studies. Research on social work practice, 23(3):266–283. Team Glm, Aohan Zeng, Bin Xu, Bowen Wang, Chen- hui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, and 1 others. 2024. Chatglm: A family of large language models from glm-130b to glm-4 all tools. arXiv preprint arXiv:2406.12793. Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations. Yuxin Hu, Danni Liu, Bo Liu, Yida Chen, Jiuxin Cao, and Yan Liu. 2025. PsyAdvisor: A plug-and-play strategy advice planner with proactive questioning in psychological conversations. In Proceedings of the 63rd Annual Meeting of the Association for Compu- tational Linguistics (Volume 1: Long Papers), pages 12205–12229, Vienna, Austria. Association for Com- putational Linguistics. Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. 2023. Beavertails: To- wards improved safety alignment of llm via a human- preference dataset. Advances in Neural Information Processing Systems, 36:24678–24704. XiaoWen Ji, Cheryl HK Chui, Shi Guang Ni, and Rui Dong. 2020. Life satisfaction of rural migrant work- ers in urban china: The roles of community service participation and identity integration. Journal of So- cial Service Research. Yukyung Lee, JoongHoon Kim, Jaehee Kim, Hyowon Cho, Jaewook Kang, Pilsung Kang, and Najoung Kim. 2025. CheckEval: A reliable LLM-as-a-judge framework for evaluating text generation using check- lists. In Proceedings of the 2025 Conference on Em- pirical Methods in Natural Language Processing, pages 15771–15798, Suzhou, China. Association for Computational Linguistics. Chin-Yew Lin. 2004. ROUGE: A package for auto- matic evaluation of summaries. In Text Summariza- tion Branches Out, pages 74–81, Barcelona, Spain. Association for Computational Linguistics. 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 3469–3483, Online. Association for Computa- tional Linguistics. Ilya Loshchilov and Frank Hutter. 2017.Decou- pled weight decay regularization. arXiv preprint arXiv:1711.05101. Mounica Maddela, Megan Ung, Jing Xu, Andrea Madotto, Heather Foran, and Y-Lan Boureau. 2023. Training models to generate, recognize, and reframe unhelpful thoughts. In Proceedings of the 61st An- nual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13641– 13660, Toronto, Canada. Association for Computa- tional Linguistics. William R Miller and Stephen Rollnick. 2012. Motiva- tional interviewing: Helping people change. Guil- ford press. Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computa- tional Linguistics, pages 311–318. Huachuan Qiu, Hongliang He, Shuai Zhang, Anqi Li, and Zhenzhong Lan. 2024. SMILE: Single-turn to multi-turn inclusive language expansion via ChatGPT for mental health support. In Findings of the Associ- ation for Computational Linguistics: EMNLP 2024, pages 615–636, Miami, Florida, USA. Association for Computational Linguistics. Huachuan Qiu and Zhenzhong Lan. 2025. PsyDial: A large-scale long-term conversational dataset for mental health support. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 21624– 21655, Vienna, Austria. Association for Computa- tional Linguistics. Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. In Advances in Neural Information Processing Systems, volume 36, pages 53728–53741. Curran Associates, Inc. Martin EP Seligman and Mihaly Csikszentmihalyi. 2014. Positive psychology: An introduction. In Flow and the foundations of positive psychology: The collected works of Mihaly Csikszentmihalyi, pages 279–298. Springer. Ashish Sharma, Kevin Rushton, Inna Lin, David Wad- den, Khendra Lucas, Adam Miner, Theresa Nguyen, and Tim Althoff. 2023. Cognitive reframing of nega- tive thoughts through human-language model inter- action. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 9977–10000, Toronto, Canada. Association for Computational Linguistics. Hao Sun, Zhenru Lin, Chujie Zheng, Siyang Liu, and Minlie Huang. 2021. PsyQA: A Chinese dataset for generating long counseling text for mental health support. In Findings of the Association for Com- putational Linguistics: ACL-IJCNLP 2021, pages 1489–1503, Online. Association for Computational Linguistics. Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, and 197 others. 2025. Gemma 3 technical report. Preprint, arXiv:2503.19786. 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 sup- port conversation. In Proceedings of the 60th Annual Meeting of the Association for Computational Lin- guistics (Volume 1: Long Papers), pages 308–319, Dublin, Ireland. Association for Computational Lin- guistics. Zhilin Wang, Yi Dong, Jiaqi Zeng, Virginia Adams, Makesh Narsimhan Sreedhar, Daniel Egert, Olivier Delalleau, Jane Scowcroft, Neel Kant, Aidan Swope, and Oleksii Kuchaiev. 2024.HelpSteer: Multi- attribute helpfulness dataset for SteerLM. In Pro- ceedings of the 2024 Conference of the North Amer- ican Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 3371–3384, Mexico City, Mexico. Association for Computational Linguistics. Mengxi Xiao, Qianqian Xie, Ziyan Kuang, Zhicheng Liu, Kailai Yang, Min Peng, Weiguang Han, and Jimin Huang. 2024. HealMe: Harnessing cognitive reframing in large language models for psychother- apy. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Vol- ume 1: Long Papers), pages 1707–1725, Bangkok, Thailand. Association for Computational Linguistics. 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. Alex Young, Bei Chen, Chao Li, Chengen Huang, Ge Zhang, Guanwei Zhang, Guoyin Wang, Heng Li, Jiangcheng Zhu, Jianqun Chen, and 1 others. 2024. Yi: Open foundation models by 01. ai. arXiv preprint arXiv:2403.04652. Chao Zhang, Xin Shi, Xueqiao Zhang, Yifan Zhu, Yi Yang, and Yawei Luo. 2025. Decoupledesc: En- hancing emotional support generation via strategy- response decoupled preference optimization. arXiv preprint arXiv:2505.16995. Chenhao Zhang, Renhao Li, Minghuan Tan, Min Yang, Jingwei Zhu, Di Yang, Jiahao Zhao, Guancheng Ye, Chengming Li, and Xiping Hu. 2024. CPsyCoun: A report-based multi-turn dialogue reconstruction and evaluation framework for Chinese psychologi- cal counseling. In Findings of the Association for Computational Linguistics: ACL 2024, pages 13947– 13966, Bangkok, Thailand. Association for Compu- tational Linguistics. Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. 2019. Bertscore: Eval- uating text generation with bert. arXiv preprint arXiv:1904.09675. Xiantao Zhang. 2025. Auradial: A large-scale human- centric dialogue dataset for chinese ai psychological counseling. In Findings of the Association for Com- putational Linguistics: EMNLP 2025, pages 2847– 2863. Haiquan Zhao, Lingyu Li, Shisong Chen, Shuqi Kong, Jiaan Wang, Kexin Huang, Tianle Gu, Yixu Wang, Jian Wang, Liang Dandan, Zhixu Li, Yan Teng, Yanghua Xiao, and Yingchun Wang. 2024. ESC-eval: Evaluating emotion support conversations in large language models. In Proceedings of the 2024 Confer- ence on Empirical Methods in Natural Language Pro- cessing, pages 15785–15810, Miami, Florida, USA. Association for Computational Linguistics. Weixiang Zhao, Xingyu Sui, Xinyang Han, Yang Deng, Yulin Hu, Jiahe Guo, Libo Qin, Qianyun Du, Shijin Wang, Yanyan Zhao, and 1 others. 2025a. Chain of strategy optimization makes large language models better emotional supporter. arXiv preprint arXiv:2503.05362. Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, and 1 others. 2025b. Swift: a scalable lightweight infrastructure for fine-tuning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 29733–29735. Chujie Zheng, Sahand Sabour, Jiaxin Wen, Zheng Zhang, and Minlie Huang. 2023. AugESC: Dialogue augmentation with large language models for emo- tional support conversation. In Findings of the As- sociation for Computational Linguistics: ACL 2023, pages 1552–1568, Toronto, Canada. Association for Computational Linguistics. Hao Zhou, Minlie Huang, Tianyang Zhang, Xiaoyan Zhu, and Bing Liu. 2018. Emotional chatting ma- chine: Emotional conversation generation with inter- nal and external memory. In Proceedings of the AAAI conference on artificial intelligence, volume 32. Jinfeng Zhou, Zhuang Chen, Bo Wang, and Minlie Huang. 2023. Facilitating multi-turn emotional sup- port conversation with positive emotion elicitation: A reinforcement learning approach. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1714–1729, Toronto, Canada. Association for Computational Linguistics. Caleb Ziems, Minzhi Li, Anthony Zhang, and Diyi Yang. 2022. Inducing positive perspectives with text reframing. In Proceedings of the 60th Annual Meet- ing of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3682–3700, Dublin, Ireland. Association for Computational Linguistics. A Additional Dataset Construction Details This appendix provides supplementary details omit- ted from the main paper due to space limits. We organize the material into four parts: dataset con- struction details, implementation and evaluation details, pilot user-study details, and qualitative ex- amples. A.1 Topic Taxonomy and Filtering StageWell is organized around ten common support topics. During preprocessing, source dialogues are normalized into a student-facing support setting when the underlying concern can be mapped to a plausible school, family, peer, or growth scenario. Dialogues with no reasonable mapping to the target setting, or with content outside the non-clinical sup- port scope, are removed. Table 8 summarizes the final SFT training-instance distribution, while Fig- ure 5 visualizes thematic and contextual coverage across the 12,445 SFT training instances. A.2 SFT Rewriting Pipeline The SFT rewriting pipeline uses a stage-aware skeleton. Each assistant turn is assigned a dialogue tool, a required information slot, and a support stage. This mapping makes the generation pro- cess auditable: the Writer realizes the planned tool, while Patcher and Refiner can change wording but not the assigned function. Table 9 gives the com- plete mapping used in the construction pipeline. To provide a complete view of the generation constraints, Figure 6 illustrates the full prompt tem- plate used by the planner-guided rewriter. This prompt strictly enforces the stage control, con- strains the scenario to student-appropriate settings, and dictates a structured JSON output to ensure the resulting skeleton remains easily parseable. A.2.1 Module Contracts and Deterministic Guards The full SFT construction workflow uses a heuristic Case Card to summarize topic, user goal, emotion, and risk; a Planner to emit the target rounds, stage schedule, fact list, and turn skeleton; a Writer to re- alize the skeleton as a complete dialogue; and Json- Fix, Patcher, and Refiner as post-process modules for parse recovery, local repair, and fluency polish- ing. In addition to these generative modules, deter- ministic Validator and Repeat Gate checks decide whether the sample is accepted, locally patched, or discarded. Table 10 summarizes the role of each module. The Planner prompt enforces several non- negotiable constraints that are especially important for corpus reproducibility. First, topic assignment is fixed by the Case Card and cannot be freely reinterpreted by the model. Second, adult or work- place scenarios must be translated into plausible school, family, or peer contexts while preserving the same practical concern. Third, the support pro- cess must remain monotonic: later stages may not appear before the earlier process has been suffi- ciently grounded. Fourth, the Planner must emit a complete JSON schema rather than free text, so that subsequent modules can operate over named fields instead of brittle string parsing. At the sample-audit level, we retain enough in- termediate structure to explain why a rewritten dialogue looks the way it does, without storing every transient decoding artifact in the paper ap- pendix. Concretely, the stored trail includes: the normalized Case Card, the final Planner schedule and fact set, the Writer dialogue with step labels, any Patcher-triggered local edits, and the final Val- idator decision. This design also supports failure re- covery: malformed JSON can be repaired without changing semantics, while true semantic violations are filtered rather than silently rewritten. Table 12 gives a compact example of the audit trail retained for one rewritten dialogue, including Planner fields, normalized facts, skeleton-level in- tent, a short rewritten-dialogue excerpt, and down- stream validation signals. In addition to the scale and stage distribution re- ported in the main paper, we track several structural diagnostics to ensure that the rewritten dialogues remain well-formed. Table 13 summarizes these checks. A.3 GroundTruth Revision GroundTruth is constructed through expert revision of automatically rewritten dialogue drafts. The ini- tial drafts are produced by the preceding rewrit- ing pipeline, and experts revise assistant responses while preserving the topic, turn structure, dialogue context, and non-clinical task scope. The main revision targets are expression adjustment, empa- thy optimization, removal of inappropriate word- ing, and replacing generic suggestions with more concrete school-scenario expressions. The final GroundTruth subset contains 120 expert-revised dialogues and 977 QA pairs. Figure 5: Thematic and contextual coverage of the SFT corpus. The inner ring shows high-level support themes, the outer ring shows representative contextual scenarios, and the side legend reports theme volumes across the 12,445 SFT training instances. A total of 28 experts with psychology or education-related backgrounds participated in for- mal revision. They were recruited through a sign- up and screening process and came from front- line educational settings, covering primary school, junior high school, senior high school, and sec- ondary vocational contexts. Their expertise in- cludes mental-health education, student develop- ment guidance, subject teaching, class manage- ment, moral education, and school-based teach- ing and research; some also had experience with student counseling, home-school communication, or growth support. This expert composition was intended to improve the educational-scenario fit, psychological appropriateness, and naturalness of the revised dialogues. Revision was organized through a web-based task system. Dialogues were assigned as task units; TopicCountProp.Typical coverage Peer relationships and campus social life 4,827 38.8% Friend conflict, exclusion, bullying, group pressure, social distance. Academic pressure and learning strategies 2,474 19.9%Homework load, exams, rank anxiety, review planning, teacher inter- action. Family communication and support1,818 14.6%Parent-child communication, parental expectations, misunderstanding, family support. Goal and strength development7686.2% Goal setting, interests, strengths, self-motivation, future direction. Emotion and stress management6235.0% General anxiety, sadness, helplessness, emotional pressure. Self-concept and self-esteem5414.3% Low confidence, self-doubt, appearance concerns, comparison, self- worth. Digital life and media use4503.6%Phone dependence, games, short videos, online conflict, information overload. Self-management and habits4153.3% Procrastination, sleep, lateness, daily planning, habit regulation. Adaptation and developmental transition 2802.2% Transfer, class change, new environment, puberty-related adaptation. Safety and boundaries2492.0% Privacy, harassment, threats, refusal difficulty, personal boundaries. Table 8: Topic taxonomy and final SFT topic distribution. Counts sum to 12,445. Figure 6: SFT prompt template used in planner-guided whole-dialogue rewriting. experts could view only the samples assigned to them and revised the rewritten drafts directly in the system. Administrators handled account man- agement, task assignment, progress tracking, and data export, but did not rewrite the dialogue con- tent. Each finalized sample therefore reflects expert editing and confirmation rather than majority-vote labeling. Since GroundTruth is an expert-revised reference set rather than a multi-rater annotation set, we do not report inter-annotator agreement; in- stead, quality control focuses on preserving context, stage intent, role boundaries, and student-facing language during expert revision and final format- ting. B Additional Implementation and Evaluation Details B.1 Full Training Hyperparameters Table 15 reports the full hyperparameter config- uration used for reproducibility. Compared with SFT, the DPO stage uses a more conservative setup, including a smaller learning rate (5× 10 −7 ) and larger gradient accumulation, to stabilize prefer- ence alignment and avoid drifting too far from the SFT distribution. Training and evaluation are im- plemented with ModelScope Swift (ms-swift) on a single-node server with2×NVIDIA L20 (48GB) GPUs and CUDA 12.8. ToolRequired slotStagePurpose check_in main_painS1Confirm the main concern and emotional load. trigger_scene recent_sceneS2Elicit a recent concrete triggering scene. body_signal body_signalS2Add bodily or pressure-related cues when explicitly grounded. self_talk inner_talkS2Clarify the user’s current thought or self-talk. stakes worst_caseS2Clarify feared consequences or perceived stakes. strength_exception past_successS3Retrieve past coping experiences or exceptions. support_system supportS3Identify available support sources. best_self_scene hope_imageS4Construct a concrete image of a better state. micro_step first_actionS5Produce a low-burden first action. if_then_trigger trigger_planS5Form an if–then trigger plan. plan_b backup_planS5Prepare a backup action when the first plan is blocked. summary_transfer summary_takeawayS6Summarize and transfer one takeaway to future use. Table 9: Dialogue-tool and information-slot mapping used in the stage-aware rewriting skeleton. ModulePrimary outputCore responsibilityHard boundary Case Card (heuristic)topic, tags, emotion, goal, risk Build a stable student-facing scenario anchor before generation. High-risk cases are routed out before ordinary rewrit- ing. Planner target_rounds, step_schedule, facts, skeleton, close_hint Convert the source dialogue into an ordered six-stage plan with evidence-linked turn functions. Must preserve topic class, monotonic stage order, and adolescent-scenario normalization. Writerrewrite[] with step labelsGenerate the whole rewritten dialogue under the shared skeleton. Cannot change stage assignment, round count, or required slots. JsonFixrepaired JSON or explicit failure mark Recover truncated or malformed structured outputs with minimal edits. May repair format only; no semantic rewriting. Patcherlocally rewritten marked rounds Repair repetition, near-duplicate questioning, or tran- sition roughness. Must keep the same step, facts, and required slot. Refinerpolished rewrite[] Improve naturalness and coherence at the discourse level. Cannot add/delete turns or drift into adult/clinical framing. Validator (rule-based) pass/fail + reason Check role alternation, end-state constraints, banned terms, and stage monotonicity. Rejected samples do not enter the final corpus. Repeat Gate (rule-based) patch trigger for flagged rounds Detect assistant-side near-repetition and invoke local repair only when needed. Cannot rewrite content by itself; it only triggers Patcher. Table 10: Generative and deterministic modules in the full SFT rewriting workflow. Planner fieldOperational meaning target_roundsDesired dialogue length after rewriting, constrained to be no shorter than the original usable interaction. step_scheduleOrdered six-stage route indicating which support stage each assistant turn should serve. close_hint Final-turn closing intention used to keep the S6 ending brief and non-interrogative. factsNormalized situation facts that downstream modules may rely on; these are adolescent-adapted and topic-aligned. skeleton[t].q_typeDialogue tool assigned to assistant turn t, e.g., trigger_scene or micro_step. skeleton[t].must_add_slot Required information slot that must appear in turntif the tool is realized correctly. skeleton[t].required_fact_idsMinimal fact set that the turn is allowed to draw on, used to reduce hallucinated detail. skeleton[t].evidence_roundSource round most directly motivating the planned assistant action, retained for auditability. Table 11: Planner-side structured fields retained in the rewriting skeleton. Audit artifactIllustrative retained content Case CardTopic: academic pressure; emotion: frustration + shame; user goal: reduce avoidance and regain a manageable first step; risk: ordinary. Planner fields target_rounds=8;step_schedule=[S1,S2,S2,S3,S4,S5,S5,S6];close_hint: briefly consolidate the student’s manageable next step. Retained factsF1: the student feels unprepared for an upcoming exam; F2: panic increases when thinking about unknown questions; F3: previous preparation routines helped; F4: teacher feedback is available as support. Skeleton tool mapping Early turns clarify the concrete trigger and self-talk; middle turns activate prior strengths and a preferred future image; later turns construct a micro-step and summarize one transferable takeaway. Rewritten dialogue excerptUser: “I feel blank whenever I think about the exam.” Assistant: validates the pressure, asks for a recent trigger, retrieves a prior coping exception, proposes a 15-minute review step, and ends with a brief non-interrogative transfer summary. Validator decisionAccepted: role alternation holds, stages are monotonic, no unsupported facts are introduced, and the final assistant turn is non- interrogative. H-score auditNo hallucination or boundary violation; evidence field highlights short assistant spans that stay within non-clinical support language. Q-targeted rewrite cueTarget dimension: Q3 feasibility. Rewrite instruction: compress advice into one low-burden action the student can try immediately after class. Table 12: Illustrative audit trail showing the kinds of intermediate artifacts retained by the construction and evaluation pipeline. This table is schematic rather than a verbatim sample dump. GroupDiagnosticValue ScaleSFT instances12,445 LengthAvg. turns per sample8.09 LengthAvg. characters per sample617.51 LengthAvg. user-turn characters24.38 LengthAvg. assistant-turn characters51.96 LengthAssistant/user length ratio2.13 StructureSamples starting with user100.00% StructureSamples ending with assistant100.00% Stage controlTurn-level consistency100.00% Round controlAvg. round increase+1.26 Safety routingHigh-risk samples0.61% Table 13: Additional structural diagnostics for the SFT subset. B.2 Data Preprocessing The complete construction and preprocessing pipeline for the SFT and DPO training corpora is described in detail in the preceding sections of this appendix. The preprocessing mainly in- cludes: structural parsing and role normalization of the source dialogues (client/positive psychology coach), removal of samples with insufficient turns (n < 7) and content clearly deviating from the ado- lescent context (e.g., mortgages, stock trading, and other adult-oriented topics), Case Card construc- tion and scenario determination, and structured rewriting via the Planner–Writer–Patch–Refiner Revision dimensionWhat experts were asked to improve Context fidelityPreserve the original topic, user concern, and situational facts unless the draft contains explicit distortion. Stage intentKeep the intended support function of each assistant turn rather than freely rewriting the dialogue strategy. Educational fitReplace adult, clinical, or over-formal wording with school-appropriate and adolescent- facing language. Empathic adequacyStrengthen emotional acknowledgment when the draft moves too fast into explanation or advice. Action realism Replace generic suggestions with lower-burden, more concrete, and context-grounded next steps. Boundary safetyRemove role overreach, diagnosis-like wording, or advice inconsistent with non-clinical support. Chinese naturalnessPolish stiff, templated, or translated-sounding Chinese into more life-like conversational phrasing. Table 14: Main dimensions emphasized during expert revision of GroundTruth drafts. HyperparameterSFTDPO InitializationPublic backboneBackbone + LoRA Frameworkms-swiftms-swift TaskLoRA SFTLoRA DPO PrecisionBF16BF16 Max seq length10241536 OptimizerAdamWAdamW Learning rate5× 10 −5 5× 10 −7 Schedulercosinecosine Warmup ratio0.050.05 Weight decay0.10.01 Train batch size11 Grad accum816 Epochs31 LoRA r / α8 / 328 / 32 DPO β–0.7 Table 15: Training hyperparameters for SFT and DPO. pipeline. All finalized samples undergo determinis- tic validation (turn alternation, step monotonicity, final-turn closure, etc.) to ensure format and struc- tural consistency across the training data. B.3 Validation Split For the SFT stage, a validation set is randomly sam- pled from the SFT corpus at a ratio of 0.01. This validation set is excluded from training and used solely for metric monitoring during the training process. No validation set is held out for the DPO stage. B.4 Model Saving and Selection The checkpoint saving and model selection strategy used during training is as follows: • Checkpoint saving: For SFT, checkpoints are saved every 100 steps, with a maximum of 2 retained. For DPO, checkpoints are saved ev- ery 20 steps, with a maximum of 10 retained. •Model selection: For SFT, the checkpoint with the lowest validation loss is selected as the final model. For DPO, the optimal check- point is determined by jointly monitoring the training loss curve and evaluation metric. If the loss continues to decrease while the eval- uation metric fails to improve in later stages, priority is given to the checkpoint with the highest evaluation metric. B.5 Hyperparameter Configuration The best-found hyperparameter values for both the SFT and DPO stages are listed in Table 15. The four base models (GLM-4-9B, Qwen3-14B, Gemma-3-12B, and Yi-1.5-9B) share identical pri- mary training hyperparameters in both stages to ensure comparability across models. B.6 Evaluation Protocol Evaluating multi-turn psychological support re- quires moving beyond general text-quality metrics to capture process-sensitive support quality. Ta- ble 16 details the five Q dimensions and their ap- plicable stages. Non-applicable dimensions are recorded as null and excluded from the turn-level mean. To reduce variance in LLM-as-a-Judge eval- uations, Table 17 provides explicit 5-point anchors for these dimensions. Table 20 then illustrates how these criteria translate into process-localized re- Dim.Core judgmentStages Q1 Content fitAddresses explicit concern without topic drift or over-reading.S1-S6 Q2 AcknowledgmentCatches user’s emotion and vulnerability before advice.S1-S6 Q3 FeasibilityRealistic, concrete, low-burden, immediately actionable.S5 Q4 MaturityNatural, proportionate, experientially plausible.S5,S6 Q5 Chinese naturalnessFluent, conversational, non-template-like Chinese.S1-S6 Table 16: Q-dimension definitions and stage applicability within the six-stage support process. pairs. The H module is applied at the dialogue level and flags factual hallucination, clinical or role- boundary violation, safety risk, and other critical boundary failures. The Q module is applied at the turn level and scores only the dimensions applica- ble to the expected support stage. The S module is a turn-level stage classifier that predicts the pri- mary and secondary six-stage labels. We compute S-exact, Top-2, Within-1, and Macro-R from these predictions. JSON Repair is used only for format recovery and does not modify content judgments. The evaluation prompts are fixed across all model conditions and require strict JSON outputs with short evidence fields. The Quality Judge does not reassign the target support stage; instead, the script passes the expected stage and masks non- applicable Q dimensions before aggregation. This design keeps Q scoring focused on local response quality while the S module separately evaluates stage recognition and progression. We use the same protocol for Base, SFT, and DPO outputs, so the re- ported differences reflect model-condition changes under a shared diagnostic procedure. B.6.1 Judge-Module Contracts The LLM-as-a-Judge stack is summarized here at the module level rather than line by line. Table 18 reports the operational contract of each module. The main design principle is separation of con- cerns: H performs dialogue-level red-line screen- ing, Q performs stage-conditioned local quality scoring, S performs stage recognition, and JSON Repair is activated only when format parsing fails. The Quality Judge includes many negative re- minders designed to stop the evaluator from over- rewarding long or polished but generic replies. In particular, the judge is told not to equate “complete” with “high quality,” not to ignore Q1 over-reading merely because the tone is warm, and not to award high Q5 scores to bookish or checklist-like Chi- nese that is understandable but still unlike natural conversation. During DPO preference construction, we also analyze the distribution of localized defects flagged by HQS, as summarized in Table 23. Because pref- erence optimization can be affected by response- length bias, we additionally examine length align- ment across stages (Table 24). The results show that chosen responses are often no longer than re- jected responses, suggesting that the DPO objective is not simply rewarding verbosity. B.6.2 Targeted DPO Rewrite Policy The prompt used for DPO chosen-response con- struction follows an intentionally narrow repair principle: the rewrite agent is told to improve only the designated target dimensionQ k while preserv- ing the same dialogue context, stage function, and core semantics. This keeps the preference signal interpretable and avoids turning a single pair into a wholesale response rewrite. Table 21 summarizes the dimension-specific rewrite rules. This prompt design is important for interpret- ing the resulting preference data. If a chosen re- sponse were allowed to freely improve every possi- ble defect, then the learned preference could blur together stage correction, tone refinement, added specificity, and length increases. By contrast, the targeted prompt approximates a controlled inter- vention: it asks the model to fix one dimension, to avoid visible collateral drift on the others, and to return a single candidate that can be compared directly with the original response under the same dialogue state. B.7 Additional Result Visualization While the main paper reports exact numerical re- sults in tabular form, Figure 9 provides complemen- tary visual comparisons for the key reference-based and HQS metrics across the Base, SFT, and DPO conditions. B.8 Artifact Compliance All pre-existing artifacts used in this work were employed in accordance with their stated terms of use and intended purposes. The open-source back- bone models—Qwen3-14B, GLM-4-9B, Yi-1.5- 9B, and Gemma-3-12B— are each released under their respective open-source licenses (Apache 2.0 Dim.5-point anchor3-point anchor1-point anchor Q1Tightly addresses the current user turn; no topic drift or unsupported inference. Broadly relevant but misses the specific focus or includes mild over-reading. Off-topic, misunderstanding, or treats unstated con- tent as fact. Q2Catches the user’s emotion and vulnerability with supportive tone. Shows acknowledgment but remains surface-level.Cold, judgmental, didactic, or moves to advice be- fore acknowledgment. Q3Gives a concrete, realistic, low-burden next step. Gives a direction but is vague, high-burden, or hard to execute. Abstract, idealized, or leaves the user unsure what to do next. Q4 Advice is natural, steady, proportionate, and life- like. Direction is reasonable but stiff, generic, or lacking lived-situation nuance. Mechanical, scripted, over-designed, or forced. Q5Fluent, natural, and conversational Chinese.Understandable but bookish, template-like, or mildly awkward. Stilted, unnatural, overly ornate, or AI-like. Table 17: Q-score anchors used by the LLM-as-a-Judge evaluator. Intermediate scores 2 and 4 are used for cases between adjacent anchors. Judge modulePrimary JSON fieldsWhat the prompt asks it to doWhat the prompt explicitly forbids Hard Judge (H)flags, alignment_score, evidence, notes Inspect the whole dialogue for hallucination, safety risk, role violation, and boundary overreach; provide short evidence spans. No free-form essay; no stage scoring; no evidence invented beyond assistant text spans. Quality Judge (Q)applicable_dims, scores, dimension_notes, evidence_by_dim, confidence Score only the candidate assistant turn on the Q di- mensions valid for the expected stage. Must not respecify the stage, score historical turns, or rate non-applicable dimensions. Stage Judge (S)pred_step_name, confidence, secondary_step_name Predict the dominant support stage realized by the current assistant turn. Must output one primary stage only; no broad expla- nation beyond the schema. JSON Repairrepaired JSON or failure object Repair malformed judge outputs into valid JSON with minimal editing. Cannot revise the substantive judgment or add new analytical content. Table 18: Operational contracts of the four modules in the LLM-as-a-Judge stack. or equivalent, with Gemma-3 under the Gemma li- cense), and were used solely for research purposes consistent with those licenses. ModelScope Swift (ms-swift) is also distributed under Apache 2.0. The source dialogues used as rewriting seeds were obtained from publicly available datasets intended Figure 7: HQS judge prompt and output schema. Schema rulePurpose in evaluation Evidence spans must be shortKeeps judgments auditable and reduces post-hoc rationalization. Non-applicable Q dimensions are null Prevents the evaluator from fabricating relevance for stages where a dimension is not meaningful. Confidence is retained separatelyAllows downstream scripts to inspect uncertain stage predictions without changing the main score definition. JSON Repair is content-neutralEnsures parsing recovery does not inflate or alter model performance. Candidate reply is isolatedEncourages true local scoring instead of diffuse holistic impressions. Table 19: Judge-side schema conventions motivated by auditability and stability. TargetUser turnRejected responseChosen response Q1 content fitA classmate knocked over my water and half my workbook got wet. I shouted and the whole class looked at me. You must have been tense; maybe your palms got sweaty and your chest felt tight. The water spilled over half your workbook, you shouted, and everyone looked at you. That moment could easily make you explode. Q2 acknowledgment I studied all week, but I still did badly on the exam. I really do not want to talk to anyone now. First make a wrong-question list and reset your study plan for the next exam. You worked hard all week and still got this result, so it makes sense that you feel upset and do not want to talk right now. Q3 feasibility If my dad reaches for my homework again, I will touch the eraser in my pocket. That little eraser is quietly saying, “I hear myself, and I am here.” Next time he reaches over, touch the eraser in your pocket for three seconds. You do not need to say anything; just do those three seconds first. Q4 maturityOkay, I will try.Next time you hear gossip, lightly touch your earlobe like pressing a pause button. Next time you hear that kind of gossip, first bring your attention back to what you are doing, then decide whether you want to re- spond. Q5 Chinese naturalness When I play basketball and get blocked, I do not get angry and can still laugh. When playing basketball, is your whole body loose or tight? When you get blocked in basketball, you ac- tually do not feel that tense, right? Table 20: Illustrative process-localized repair examples. Rejected responses exhibit the localized defect targeted by the Q dimension, and chosen responses repair it under the same context and stage constraint. TargetPrimary rewrite instructionWhat mustnot happen during repair Q1Re-align the reply to the most explicit content of the current user turn; remove unsupported detail completion and topic drift. Do not add inferred body reactions, inner speech, hidden motives, or new situational facts. Q2Strengthen emotional acknowledgment before advancing the task; make the user feel “received” first. Do not become preachy, overly explanatory, teasing, or theatrically gentle. Q3Compress the advice into a concrete, low-burden, immediately doable next step that fits the current scene. Do not expand into a checklist, generic self-help package, or abstract encouragement. Q4Keep the intended direction but rewrite it as something a mature, believ- able adult would naturally say in conversation. Do not rely on symbolic rituals, over-designed metaphors, script-like mini exercises, or stock reassurance. Q5Preserve meaning while rewriting the sentence into plain, natural, conver- sational Chinese. Do not pursue literary freshness at the cost of naturalness; remove “AI- like” wording and stiff template phrasing. Table 21: Dimension-specific rules for the targeted DPO rewrite prompt. Each chosen response is intended to repair one localized defect rather than globally rewrite the entire answer. Shared prompt ruleReason Fix only one target dimension at a time Prevents preference ambiguity and preserves localized supervision. Keep reply length close to the original Reduces accidental length preference and verbosity bias. Preserve stage functionEnsures DPO does not learn a different support step than intended. Avoid new safety or factual risksStops local repairs from introducing boundary regressions. Output only the rewritten assistant turn Simplifies export into chosen/rejected training format. Table 22: Shared constraints in the DPO rewrite prompt regardless of target Q dimension. DiagnosticCountProp. Single low-quality dimension72939.43% Two low-quality dimensions884.76% Three low-quality dimensions1,03255.81% Most common combo: Q1,Q2,Q576241.21% Chosen shorter than rejected98953.49% Table 23: Additional diagnostics for DPO preference pairs. StageChosen shorter S1 Rapport & Safety54.66% S2 Clarify & Reflect55.56% S3 Strengths & Resources51.63% S4 Future & Hope61.19% S5 Goals & Strategies≈46.0% S6 Summary & Transfer59.02% Table 24: Length-alignment diagnostic by target stage. Figure 8: Abridged prompt card for targeted DPO localized repair. The rewrite keeps the dialogue context and target stage fixed while repairing only the designated Q dimension. BaseSFTDPO (a) BERTScore 0.2 0.24 0.28 0.32 0.292 0.295 0.314 Qwen3 0.263 0.291 0.295 GLM-4 0.282 0.286 0.289 Yi-1.5 0.211 0.291 0.296 Gemma-3 (b) Q Overall 1 2 3 4 3.56 3.97 4.23 Qwen3 3.30 3.79 3.94 GLM-4 2.38 3.77 3.92 Yi-1.5 1.44 3.75 3.88 Gemma-3 (c) H-critical Ratelower is better 0 0.2 0.4 0.6 0.058 0.050 0.033 Qwen3 0.0330.033 0.042 GLM-4 0.142 0.042 0.033 Yi-1.5 0.575 0.0330.033 Gemma-3 (d) S-exact Accuracy 0.4 0.5 0.6 0.7 0.419 0.476 0.515 Qwen3 0.402 0.680 0.692 GLM-4 0.417 0.724 0.726 Yi-1.5 0.459 0.703 0.709 Gemma-3 Figure 9: Additional visualizations of reference-based and process-aware evaluation results across Base, SFT, and DPO conditions. for academic research, and all resulting StageWell data were filtered to exclude content outside the intended non-clinical support scope. C Feasibility and Usability Pilot Details C.1 Procedure The focused pilot includes three stages: pre-test, free interaction, and post-test. Participants first complete an immediate-state questionnaire, then interact with the deployed Qwen3-14B DPO sys- tem around everyday concerns such as academic pressure, peer relationships, or self-evaluation. The post-test repeats the immediate-state questionnaire and adds a subjective-experience questionnaire. Ta- ble 25 summarizes the protocol. The design is descriptive and uncontrolled, and is intended to characterize short-term user experience rather than estimate clinical or causal effectiveness. C.2 Participant Information, Consent, and Ethics Before the pilot, participants and their guardians were provided with a written participant- information and informed-consent form in Chinese. We summarize the participant-facing instructions, recruitment setting, consent process, privacy protection, and ethics-related information below. Figure 10 provides a compact visual rendering of the consent summary. Participant-facing instructions. The participant- and-guardian information sheet stated the follow- ing points in substance: the study was conducted by the research team and examined the user experi- ence, perceived support, and practical feasibility of an LLM-based positive-psychology dialogue sys- tem for adolescents in educational settings. Par- ticipants were invited to help evaluate whether the system could provide some support for stress regu- lation, emotional relief, and problem clarification in everyday adolescent growth scenarios. The tar- get participants were adolescent students. During the study, participants were asked to complete sev- eral questionnaires about their current psycholog- ical state and usage experience, and to engage in a multi-turn web-based dialogue with the system. The dialogue topics focused on common adolescent situations such as academic pressure, peer interac- tion, self-evaluation, and growth-related concerns. The full procedure was expected to take about 30 minutes. The same instruction sheet further stated that participation would not affect school evaluation, school management, or any other personal rights or interests. Participants were told to answer and ex- press themselves according to their real feelings. If they felt tired, uncomfortable, or no longer wished to continue while completing the questionnaires or interacting with the system, they could stop at any time without giving a reason and without any negative consequence. The study was described as low risk: it mainly involved everyday emotional experiences and growth-related issues and was not expected to create substantial risk, although some questions might touch on personal pressure or un- pleasant experiences and could therefore cause mild emotional fluctuation. The researchers stated that they would make efforts to keep the proce- dure appropriate and safe and would provide fur- ther explanation when necessary. The sheet also stated that the collected information would be used only for academic research; real names and other directly identifying information would not be dis- closed; results would be reported only in aggregate and anonymized form; and the materials would not be casually shared or used for unrelated purposes. Participants and guardians were asked to read the document carefully, ask questions if needed, and decide whether to participate only after fully under- standing the study. Recruitment setting and compensation. The fo- cused pilot sample consisted of junior and senior high school students from the school-side deploy- ment setting described in the main paper rather than from a crowdsourcing platform. The 28 partic- ipants included 14 junior high school students and 14 senior high school students, with an average age of 15.2 years. Recruitment was conducted offline in the school setting. No crowdsourcing platform was used, and no crowdsourcing-style payment was involved. Consent, privacy, and data use. Participation proceeded only after the participant-and-guardian information sheet had been provided and informed consent had been obtained. The form explained the study purpose, procedure, approximate dura- tion, low-risk nature, voluntary participation, with- drawal rights, and privacy protections. Responses and dialogue logs collected in the pilot study were used only for academic research, analyzed in de- identified form, and reported only at the aggregate level in this paper. Ethics review and risk control. The pilot was pre- sented to participants as a low-risk, voluntary study with explicit withdrawal rights and privacy protec- tion. The study was reviewed and approved by the authors’ institutional ethics committee. The system StageOperationPurpose Pre-testFill in immediate-state scale.Record state before interaction. Free interactionMulti-turn supportive dialogue.Provide short-term support. Post-test IFill in the same state scale.Compare pre/post state. Post-test IIFill in subjective-experience scale.Evaluate feasibility and usability. Table 25: Feasibility and usability pilot procedure. Figure 10: Reconstructed participant-information and informed-consent summary used in the pilot. Blue text: key study metadata. Red text: major consent dimensions.Green background: participant protections, intended support scope, and withdrawal rights.Yellow background : low-risk and anonymized-reporting statements. was positioned as non-clinical support for every- day growth-related concerns and not as diagnosis, treatment, or crisis intervention. C.3 Questionnaire Items All questionnaire items use a five-point Likert scale. For the immediate-state scale, the first two items are reverse-coded so that higher scores always in- dicate a more positive immediate state. In the pi- lot sample, Cronbach’s alpha is 0.871 for pre-test, 0.870 for post-test, and 0.887 when pre/post re- sponses are pooled. These internal-consistency statistics are reported only for this small pilot sam- ple; the scale is used to describe short-term within- session changes and is not a standardized clinical instrument or diagnostic measure. For the subjective-experiment analysis, letx (t) ij denote participanti’s score on itemjat time t ∈ pre, post, withm = 5immediate-state items. The first two items (stress burden and emo- tional distress) are reverse-coded; after recoding, we write the aligned score as ̃x (t) ij . The participant- level immediate-state score at time t is defined as S (t) i = 1 m m X j=1 ̃x (t) ij .(4) The individual pre–post change is then ∆ i = S (post) i − S (pre) i ,(5) where∆ i > 0indicates an improvement in imme- diate state after interaction. At the sample level, the mean change is computed as ̄ ∆ = 1 N N X i=1 ∆ i .(6) Table 26 lists the five immediate-state items used in the pre/post questionnaire. For the post- interaction subjective-experience questionnaire, let y ik denote participanti’s score on dimensionk. We DimensionItem Stress burden (rev.)I currently feel a relatively high level of pressure. Emotional distress (rev.)I am still clearly troubled by the current issue. RelaxationI feel calmer and more relaxed than at the beginning. HopeI feel that the current issue can gradually improve. Self-regulationI feel capable of coping with the current difficulty. Table 26: Immediate psychological-state items. The first two items are reverse-coded for scoring. report both the mean score ̄y k = 1 N N X i=1 y ik ,(7) and the positive-rate statistic p k = 1 N N X i=1 I(y ik ≥ 4),(8) whereI(·)is the indicator function. This matches the descriptive reporting style used in the pilot re- sults table in the main paper. Table 27 lists the post-interaction subjective- experience items. D Representative Prompt Excerpts This section presents the main prompt modules in a visual card format so that the instruction blocks, decision rules, and JSON-output constraints can be inspected more directly. D.1 Detailed Prompt Cards Figures 11–22 summarize the prompt cards used by the main construction, judging, and targeted- rewrite modules. We present them as compact visual excerpts to make the control rules, output schemas, and module boundaries inspectable with- out reproducing every implementation-specific in- struction line in the appendix. Taken together, these condensed materials show that StageWell is not merely a collection of outputs from a single prompt. Instead, it is a pipeline with explicit contracts between modules, auditable struc- tured outputs, deterministic acceptance checks, and evaluation prompts designed to localize rather than blur process-relevant defects. This is precisely the property that makes the dataset suitable for both SFT training and process-localized DPO align- ment. DimensionItem UnderstandingThe system understood the problem I was facing. AcceptanceThe system’s response made me feel accepted rather than judged. HelpfulnessThe system’s response was helpful to me. Guidance clarityThe system’s guidance was clear; I knew what I could think or do next. Continued-useIf I encounter a similar concern, I would be willing to use it again. SatisfactionOverall, I was satisfied with this experience. UsabilityThe web interface was easy to understand and operate. Table 27: Post-interaction subjective-experience items. Figure 11: Detailed prompt card for the Planner module. Figure 12: Detailed prompt card for the Writer module. Figure 13: Detailed prompt card for the Hard Judge module. Figure 14: Detailed prompt card for the Quality Judge module. Figure 15: Detailed prompt card for the Stage Judge module. Figure 16: Detailed prompt card for the JSON Repair module. Figure 17: Detailed prompt card for the Q1 targeted rewrite instruction. Figure 18: Detailed prompt card for the Q2 targeted rewrite instruction. Figure 19: Detailed prompt card for the Q3 targeted rewrite instruction. Figure 20: Detailed prompt card for the Q4 targeted rewrite instruction. Figure 21: Detailed prompt card for the Q5 targeted rewrite instruction. Figure 22: Detailed prompt card for the shared Q-targeted rewrite template.