cs.AIApr 6, 2026
SavePSY-STEP: Structuring Therapeutic Targets and Action Sequences for Proactive Counseling Dialogue Systems
Organizations: Graduate School of Artificial Intelligence, POSTECH · MILA · McGill University · Department of Computer Science and Engineering, POSTECH
Abstract
Cognitive Behavioral Therapy (CBT) aims to identify and restructure automatic negative thoughts pertaining to involuntary interpretations of events, yet existing counseling agents struggle to identify and address them in dialogue settings. To bridge this gap, we introduce STEP, a dataset that models CBT counseling by explicitly reflecting automatic thoughts alongside dynamic, action-level counseling sequences. Using this dataset, we train STEPPER, a counseling agent that proactively elicits automatic thoughts and executes cognitively grounded interventions. To further enhance both decision accuracy and empathic responsiveness, we refine STEPPER through preference learning based on simulated, synthesized counseling sessions. Extensive CBT-aligned evaluations show that STEPPER delivers more clinically grounded, coherent, and personalized counseling compared to other strong baseline models, and achieves higher counselor competence without inducing emotional disruption.
Figures & tables
Figure 1: Example of a structured CBT interaction for eliciting automatic thoughts and cognitive reframing.
Figure 2: Overview of the Psy-Step dataset construction and structured CBT counseling flow. The figure illustrates how client profiles are modeled, how surface-level problems and automatic thoughts are elicited during the diagnostic stage, and how structured action sequences guide therapeutic interventions through stepwise CBT reasoning.
| Dataset | Counseling Theory | Problem Representation | Problem Source | Intervention Structure | Open | Language | # of Dialogues | Avg. Turns |
| PsyCon Mishra et al. (2023) | Not Specified | Disorder-specific Experiences | Online Forum | None | English | 1,020 | 24.6 | |
| SmileChat Qiu et al. (2024) | Not Specified | Mental Health Questions | Online Q&A Platforms | None | Yes | Chinese | 55,165 | 10.4 |
| Psych8k Liu et al. (2023) | Cognitive Behavioral Therapy + Others | Patient-reported Concerns | Counseling Records | None | Yes | English | 8,187 | 10.0 |
| HealMe Xiao et al. (2024) | Cognitive Behavioral Therapy | Negative Thoughts | Crowdsourced | Planning | No | English | 1,300 | 3.0 |
| CBT-LLM Na (2024) | Cognitive Behavioral Therapy | Mental Health Questions | Online Q&A Platforms | Planning | No | Chinese | 22,327 | 1.0 |
| CACTUS Lee et al. (2024) | Cognitive Behavioral Therapy | Negative Thoughts | Crowdsourced | Planning | Yes | English | 31,577 | 16.6 |
Table 1: Comparison of counseling dialogue datasets across theory, structure, and problem representation.
Figure 3: Illustration of the simulation-based process for collecting preference data for utterance selection.
| Model | General Skills ( ) | CBT-specific Skills ( ) | Ques.-Ref. Diversity ( ) | |||||
| Understand | Interpers. | Collabo. | Guided Dis. | Focus | Strategy | AT. Coverage | ||
| GPT-4o | 3.74 | 5.63 | 5.04 | 3.94 | 3.73 | 2.58 | 2.49 | 1.07 |
| gemini-2.0-flash | 3.78 | 5.23 | 4.35 | 3.73 | 3.81 | 3.35 | 2.82 | 1.42 |
| SmileChat | 2.62 | 4.18 | 3.22 | 2.71 | 3.15 | 2.45 | 1.54 | 1.31 |
| CBT-LLM | 2.93 | 4.13 | 2.37 | 2.66 | 3.71 | 3.29 | 2.95 | 1.66 |
| Llama-Psych8k | 4.01 | 5.74 | 5.62 | 4.58 | 4.66 | 4.59 | 3.77 | 1.62 |
Table 2: Evaluation of counselor competence across models. Abbreviations: Interpers. = Interpersonal Effectiveness; Collabo. = Collaboration; Dis. = Discovery; AT. = Automatic Thought; Ques. = Question; Ref. = Reflection.
| Without Planning | |||||
| GPT-4o | Llama-psych8k | Stepper SFT_NoPlan | |||
| Q.identify | 75.86 | Q.identify | 54.85 | Q.identify | 27.66 |
| Q.alt | 7.76 | Q.reality | 13.73 | Q.thought | 18.07 |
| Q.evidence | 7.76 | Q.evidence | 7.35 | Q.reality | 17.45 |
| R.emotion | 57.84 | R.reframe | 39.59 | R.emotion | 43.31 |
| R.reframe | 15.18 | R.emotion | 27.04 | R.reframe | 22.85 |
Table 3: Distribution of top-3 question (Q.) and reflection (R.) action types, reported as percentages (%). See Appendix D.3 for tag definitions.
| Model | Withdrawn | Resistant | Engaged | All | ||||
| Helpful | Hindering | Helpful | Hindering | Helpful | Hindering | Helpful | Hindering | |
| GPT-4o | 3.16 | 1.83 | 3.02 | 2.26 | 3.60 | 1.58 | 3.29 | 1.86 |
| gemini-2.0-flash | 3.09 | 2.08 | 2.81 | 2.55 | 3.51 | 1.82 | 3.16 | 2.13 |
| CBT-LLM | 2.70 | 2.49 | 2.32 | 3.15 | 3.02 | 2.38 | 2.71 | 2.65 |
| SmileChat | 2.90 | 2.18 | 2.67 | 2.83 | 3.31 | 1.97 | 2.99 | 2.30 |
| Llama-psych8k | 3.28 | 1.91 | 3.12 | 2.15 | 3.74 | 1.55 | 3.41 | 1.84 |
Table 4: Session Rating Scale results across client engagement types, reporting Helpful and Hindering reactions.
| Model | Helpful Outcomes (↑) | |||
| Perceived | Empower- | Emotional | Self- | |
| Support | ment | Relief | Acceptance | |
| GPT-4o | 4.72 | 3.41 | 3.00 | 3.18 |
| gemini-2.0-flash | 4.47 | 3.16 | 2.76 | 3.07 |
| SmileChat | 4.16 | 3.13 | 2.63 | 2.82 |
| Llama-psy8k | 4.53 | 3.47 | 2.93 | 3.21 |
Table 5: Comparison of helpful and hindering counseling outcomes across models.
Figure 4: Preference comparisons of Stepper , conducted with Gemini-based clients and evaluators.
Figure 5: Preference comparison of Stepper conducted with human experts (see Appendix E.3 for details).
Figure 6: Correlation between overall human preference and individual counseling dimensions (Spearman’s ).
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
| Case 1: Guiding beyond empathic validation | |||
| Empathy-Oriented Model | Stepper SFT + Pref. | ||
| Example A: Frustration around goal achievement | |||
| Client | I understand your point, but I still feel like there’s more to this frustration than just achieving goals. | Client | I think it just makes reading more stressful for me, like I’m not just enjoying the book but also worrying about not keeping up or understanding enough. |
| Counselor | Yes, accepting your imperfections may take some time and effort, but I believe you can do it . | Counselor | That’s a valid concern. But can you think of any evidence that might contradict that belief? For example, have there been times when you’ve enjoyed a book, even if it took longer to finish? |
| Example B: Overwhelming distress and lack of direction | |||
| Client | I know you’re right, but I still just feel so overwhelmed. It feels so big and like anything I try won’t make a difference. It’s like, where do I even start? | Client | It feels like I’m constantly on edge, worried about every little thing at work. It’s exhausting and makes me question if I can really keep up. |
Table 6: Qualitative comparison of Stepper against diverse counseling baselines, including empathy-oriented ( SmileChat ), general purpose ( GPT-4o ), and CBT-based models ( Camel ). Across all cases, baseline models tend to provide abstract validation or underspecified cognitive reframing, whereas Stepper consistently engages with clients’ automatic thoughts and delivers more specific, and action-grounded guidance.
| Case 1: Effects of Planning | |||
| Stepper SFT_NoPlan | Stepper SFT + Pref. | ||
| Example A: Global self-judgment after failure | |||
| Client | I feel like I failed as a designer, and that makes it hard to even think about reaching out to new clients. | Client | I feel like this failure defines me as a designer. |
| Counselor | It makes sense that you’d feel hesitant after such a setback. We can take things slowly and explore these feelings together. | Counselor | What I’m hearing is an all-or-nothing conclusion. To work with that, I’m going to suggest a continuum approach—placing this mistake somewhere between “minor setback” and “career-ending failure” rather than at one extreme. |
| Example B: Overgeneralization from a negative aspect | |||
| Client | It just feels like everything went wrong. The trip was nice, but the money part ruined everything. | Client | I keep thinking about the money we spent, and it feels like the whole vacation was a disaster. |
Table 7: Qualitative comparisons across Stepper variants.
| Model | Insight | Perceived Support | Cognitive Dist. | Empowerment | Therapeutic Stuckness | Interpersonal Hope | Goal Clarity |
| GPT-4o | 2.98 | 4.72 | 2.31 | 3.41 | 2.49 | 2.84 | 3.61 |
| gemini-2.0-flash | 2.94 | 4.47 | 2.27 | 3.16 | 2.83 | 2.62 | 3.67 |
| CBT-LLM | 2.70 | 3.66 | 2.17 | 2.57 | 3.11 | 2.44 | 3.11 |
| SmileChat | 2.65 | 4.16 | 2.25 | 3.13 | 2.74 | 2.86 | 3.26 |
| Camel | 3.33 | 4.51 | 2.45 | 3.37 | 2.20 | 2.97 | 3.87 |
| Llama-psy8k | 3.32 | 4.53 | 2.50 | 3.47 | 2.11 | 2.96 | 3.94 |
Table 8: Session Rating Scale (SRS) results averaged across the 14 evaluation metrics. Higher scores generally indicate more positive client-reported experiences, whereas lower scores indicate better outcomes for hindering-related metrics.
| Expert Feedback H1 (MPhil, Clinical Psychology) |
| Overall Clinical Validity: In clinical practice, a distinction is often made between manualized therapy and the more dynamic process of real-world sessions. Within this context, the counselor’s use of Socratic questioning and Evidence for / Evidence against closely aligns with standard CBT training and reflects how clinicians help clients decenter from maladaptive thoughts. |
| Usefulness of Structured Design: The explicit linkage between surface problems, automatic thoughts, and counseling actions is particularly valuable. By maintaining this linkage, the dataset helps prevent therapeutic drift and supports more focused, clinically grounded interventions. |
| Practical Value: Overall, the dataset is well suited for training dialogue systems in logical and therapeutic consistency, demonstrating that effective counseling requires strategic, goal-directed intervention in addition to empathy. |
| Expert Feedback H2 (Master’s degree, Clinical Psychology) |
Table 9: Qualitative expert feedback (H1–H3) on the clinical validity and practical value of the dataset.
| Example 1 |
| Negative Thought: |
| No one really cares about me. |
| Attitude: |
| Over Compliant |
| Surface-Level Problem: |
| Feeling discouraged because people do not attend my parties. |
Table 10: Examples of decomposing client narratives into negative thoughts, surface-level problems, triggering situations, and automatic thoughts.
| Interaction Style | Description |
| Hesitant | Type: Withdrawn |
| Definition: Speaks cautiously and with reluctance; provides minimal information unless gently encouraged. | |
| Behavior Signals: Short answers; pauses before responding; expressions such as “I’m not sure…”; avoidance of direct emotional expression. | |
| Guarded | Type: Withdrawn |
| Definition: Avoids sharing personal details or emotions and minimizes the significance of concerns. | |
| Behavior Signals: Downplaying issues; statements like “It’s nothing serious…”; emotionally flat tone; vague or indirect responses. |
Table 11: Client interaction styles with corresponding engagement types, definitions, and behavioral signals.
| CBT Technique | Description |
| Efficiency Evaluation | Evaluates whether a thought is helpful or harmful in real-life situations. |
| Pie Chart Technique | Breaks down how different factors contribute to an event, reducing self-blame. |
| Alternative Perspective | Encourages considering how others might interpret the same situation. |
| Decatastrophizing | Reduces worst-case thinking by examining real likelihood and coping options. |
| Pros and Cons Analysis | Weighs the benefits and drawbacks of a specific thought or belief. |
| Evidence-Based Questioning | Examines evidence for and against the client’s thought. |
Table 12: List of CBT strategies used, adapted from Lee et al. (2024) , excluding strategies that are difficult to implement through dialogue alone..
| Example 1 |
| Surface-Level Problem: |
| I feel anxious about social gatherings. |
| Triggering Situation: |
| Thinking about attending a friend’s party. |
| Automatic Thoughts: |
| They must think I’m dull or antisocial. |
Table 13: Example plans illustrating how surface-level problems, triggering situations, and automatic thoughts are translated into structured CBT plans with ordered action sequences.
| Metric | Description |
| Evaluation Metrics for Utterance | |
| Alignment with Action | Assesses whether the utterance appropriately follows the expected therapeutic progress given the dialogue context and the planned action. |
| Validation & Warmth | Evaluates how well the utterance validates the client’s emotional experience and communicates warmth, empathy, and non-judgmental support. |
| Clarity | Assesses how clear, understandable, and accessible the utterance is for the client. |
| Evaluation Metrics for Plan and Action Sequence | |
| Completeness | Assesses how fully the plan includes the essential elements of a CBT-informed therapeutic step. |
Table 14: Evaluation metrics used to assess the quality of counseling utterances and plan–action sequences.
| Example 1 | |
| History | Client: I’m not sure. It feels like if I allow myself to feel differently, I might end up being careless with money. I feel like I need this fear to keep me in check. |
| Chosen | Action: identify specific financial fears |
| Utterance: Jane, it sounds like you’re feeling a lot of pressure around your finances. What are some specific fears you have about your financial situation ? | |
| Rejected | Action: identify specific financial fears |
| Utterance: What specific financial worries do you think would be the worst outcome for you ? | |
| Example 2 | |
Table 15: Comparison examples between chosen and rejected responses for utterance selection.
| Example 1 | |
| Surface | I feel inadequate compared to other pianists. |
| Trigger | Watching performances of skilled pianists online. |
| Automatic | “I’ll never be good enough to play like them.” |
| Chosen | Strategy : Decatastrophizing Action order: invite worst-case scenario explore fears and doubts evaluate probability of scenario discuss evidence for fears identify past successes and strengths develop coping strategies plan |
| Rejected | Strategy : Decatastrophizing Action order: restate failure belief clearly rate belief intensity explore likelihood of failure identify evidence against failure discuss alternative outcomes develop coping strategies together |
| Example 2 | |
Table 16: Examples of chosen and rejected action sequences collected through simulation-based preference generation. Surface and Automatic denote the client’s surface-level problem and automatic thought, respectively. For brevity, detailed planning rationales are omitted, and the full diagnostic dialogue used as input is not shown due to length; instead, condensed client profile information is provided.
| CTRS Metric | Description |
| Understanding | Accurately understands and reflects the client’s explicit and implicit concerns, demonstrating empathic listening and a clear grasp of the client’s internal experience. |
| Interpersonal Effectiveness | Maintains a positive therapeutic relationship through warmth, genuineness, confidence, professionalism, and appropriate interpersonal behavior. |
| Collaboration | Engages the client as an active partner in goal-setting and decision-making through respectful, adaptive, and non-confrontational collaboration. |
| Guided Discovery | Uses questioning and guided exploration to help the client gain insight and draw conclusions, rather than relying on persuasion or lecturing. |
| Focus | Identifies and maintains attention on the client’s key cognitions or behaviors that are most relevant to change. |
| Strategy | Applies a coherent and appropriate CBT strategy that effectively promotes cognitive or behavioral change. |
Table 17: CTRS-based evaluation metrics and their descriptions used to assess counseling quality.
| Tag | Description |
| CBT Question Tags | |
| Q_Evidence | Asking the client to identify evidence that supports or contradicts their automatic thoughts. |
| Q_Alternative | Asking the client to consider alternative perspectives, such as how another person might interpret the same situation. |
| Q_WorstScenario | Asking the client to articulate the worst possible outcome they fear in order to examine catastrophic expectations. |
| Q_Uility | Asking the client to evaluate how helpful or unhelpful a particular thought is in real-life contexts. |
| Q_Advantage | Asking the client to identify potential advantages or perceived benefits of maintaining a specific thought or behavior. |
Table 18: CBT-informed micro-action tags and their functional descriptions used for counselor utterance annotation.
| Metric | Question |
| Insight | I realised something new about myself or other people. |
| Perceived Support | I feel understood, supported, or reassured by my therapist. |
| Cognitive Distance | I feel more distanced from certain feelings, thoughts, or memories. |
| Empowerment | I feel more empowered, hopeful, or positive about myself. |
| Therapeutic Stuckness | (Hindering) I feel stuck, blocked, or unable to progress in therapy. |
| Interpersonal Hope | I feel more positively or hopeful about another person(s). |
Table 19: Client-reported evaluation metrics and corresponding questions used to assess session-level outcomes.