Papers
Topics
Authors
Recent
Search
2000 character limit reached

PSY-STEP: Structuring Therapeutic Targets and Action Sequences for Proactive Counseling Dialogue Systems

Published 6 Apr 2026 in cs.AI | (2604.04448v1)

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.

Summary

  • The paper demonstrates that decomposing CBT dialogues into distinct diagnostic and intervention phases significantly improves model fidelity in digital CBT delivery.
  • It presents the PSY-STEP dataset and the STEPPER model, which use a two-stage script with explicit planning and parameter-efficient adaptation to optimize therapeutic action sequences.
  • Experimental evaluations indicate that STEPPER outperforms baseline models in both counselor competence and client satisfaction metrics, validating the structured approach.

Structuring Therapeutic Targets and Action Sequences in Proactive Counseling Dialogue Systems: An Analysis of PSY-STEP

Introduction

"PSY-STEP: Structuring Therapeutic Targets and Action Sequences for Proactive Counseling Dialogue Systems" (2604.04448) addresses fundamental modeling limitations in digital cognitive behavioral therapy (CBT) dialogue agents. The work targets the failure of previous systems to operationalize clinically relevant constructs—namely, the explicit identification of automatic negative thoughts and the structured, strategy-aligned therapeutic plans central to efficacious CBT interventions. The authors present the PSY-STEP dataset, which structurally decomposes counseling dialogues to separate surface-level problem description, automatic thoughts, and ordered sequences of counselor actions, and STEPPER, a dialogue agent architecture leveraging this structure for improved CBT-aligned interaction.

Dataset Contributions and Methodology

The PSY-STEP dataset is constructed to resolve two key deficits in prior corpora: (1) entanglement of surface-presenting complaints with true cognitive therapeutic targets and (2) insufficient procedural granularity in representing intervention strategies through sequences of counselor actions. Using GPT-4O-MINI as the generative backbone, the authors induce dialogue with an explicit split between the diagnostic (problem identification and thought elicitation) and intervention phases (strategic cognitive restructuring), each governed by step-wise action plans.

Each synthetic counseling interaction is generated with a two-stage script: the first elicits the situation, surface-level complaint, and automatic thoughts; the second deploys dynamic planning to instantiate a client-tailored action sequence, operationalizing one of several core CBT strategies (e.g., decatastrophizing, evidence-based questioning). All resulting dialogues undergo stringent CTRS-based quality filtering. Post-filtering, the dataset comprises 6,425 multi-turn dialogues annotated with surface/automatic thought labels, structured plans, and action sequences.

STEPPER Model and Training

STEPPER is fine-tuned with parameter-efficient Low-Rank Adaptation (LoRA) leveraging the highly structured PSY-STEP dialogues. The system architecture distinguishes between an utterance generation adapter and a planner adapter. The utterance adapter explicitly reasons at each turn about plan progression and dynamically conditions each response on previous context and the next candidate action in the action sequence. The planner adapter constructs ordered action lists post-diagnosis for the intervention stage, informed by both scenario specifics and preselected CBT strategies.

To further optimize for plan adherence, action specificity, and empathic nuance, the system undergoes Direct Preference Optimization (DPO). Simulated client-counselor sessions (with GPT-4O-based client and evaluator) generate pairwise data for DPO, scoring responses/actions across axes of action consistency, empathy, and strategic clarity. The final dataset for preference learning contains over 26,000 paired utterance samples and more than 6,000 structured plan samples.

Experimental Evaluation

Evaluation Benchmarks

The evaluation leverages both standardized clinical and user satisfaction instruments:

  • Cognitive Therapy Rating Scale (CTRS): Used by both automatic and expert human annotators to assess clinical competence over general competencies and CBT-specific skills (guided discovery, focus, strategic application).
  • Session Rating Scale (SRS): Used for quantifying client-reported dimensions such as helpful and hindering reactions, empowerment, and therapeutic stuckness.

Comparisons are performed against strong general-purpose (GPT-4O, GEMINI-2.0-FLASH), empathy-optimized (SMILECHAT), and CBT-specific (CAMEL, LLAMA-PSYCH8K, CBT-LLM) baseline agents.

Numerical Results and Analysis

Counselor Competence:

STEPPER consistently achieves the highest scores on CTRS metrics, especially in dimensions requiring clinically valid guided reasoning—guided discovery, intervention focus, and strategic specificity. For instance, the structured version (STEPPERSFT) scores 4.77 (of 6) in understanding, 5.85 in interpersonal effectiveness, and 5.22 in automatic thought coverage. Preference-tuned STEPPER further improves empathic and plan-consistent behavior.

Client Satisfaction:

On the SRS, STEPPER variants (post preference tuning) yield maximum or near-maximum scores on helpful outcomes (insight, perceived support, empowerment, goal clarity) and minimize hindering reactions (stuckness, guidance deficit, emotional deterioration). Notably, hindering reaction scores for STEPPER SFT+PREF. are consistently below 1.7 compared to 2.1–2.7 in leading baselines.

Ablations:

Disabling explicit planning (STEPPERSFT_NOPLAN) markedly reduces clinical competence on CBT-specific skills, reflecting the necessity of structured guided action for meaningful cognitive intervention. Strategy diversity (turn-level entropy of question/reflection tags) is also significantly reduced in less-structured models, confirming model reliance on plan-driven orchestration.

Human Expert Evaluation:

Head-to-head human annotation trials confirm substantial preferences for STEPPER over both general and CBT-oriented baselines, with human raters prioritizing competence in guidance, strategy selection, and action specificity over mere interpersonal skill.

Generalization:

The system's relative performance persists when evaluated entirely with non-GPT clients and evaluators (e.g., Gemini-2.0-Flash), establishing that gains are not an artifact of GPT-based simulation.

Theoretical and Practical Implications

The formal separation of surface-level complaints and automatic thoughts, together with proceduralized intervention sequences, operationally instantiates critical elements of evidence-based CBT in an LLM-mediated counseling system. This approach enables both greater fidelity to clinical intent and finer model interpretability, facilitating future research in safety-critical human-computer interaction domains.

On the practical plane, the PSY-STEP framework not only supports research on rigorous, explainable digital health agents but also provides a reusable protocol for dialogue structure in adjacent therapeutic domains, such as exposure protocols, behavioral activation, and schema-focused approaches. Furthermore, the preference learning pipeline evidences that high-quality plan adherence and enhanced relational skills can be imparted without reliance on real-world sensitive records, achieving strong alignment with specialist judgment and simulated client satisfaction.

Limitations and Directions for Future Work

Notwithstanding the advances, human evaluation was restricted to expert simulation and constrained user styles; the full clinical validity and safety of deploying STEPPER-like agents in live, real-world therapy remains unaddressed. The exclusive reliance on synthetic data limits diversity, and emotional optimization (beyond strategic competence) received less attention. Future directions include advanced personalization, incorporation of longitudinal trajectory modeling, real patient pilot studies, and extension to blended (multistage, multimodal) interventions.

Conclusion

The PSY-STEP dataset and the STEPPER model formalize a decisive shift in LLM-based counseling research: away from undifferentiated surface support and generic empathy, toward clinically anchored models that systematically target the core processes of cognitive intervention. The results establish that plan-guided, action-sequenced agents trained on decomposed CBT dialogue outperform both state-of-the-art general purpose systems and prior domain-specialized baselines on both counselor competence and client-side metrics. This structured modeling paradigm sets a new standard for the design and evaluation of AI-powered therapeutic dialogue systems.

Paper to Video (Beta)

No one has generated a video about this paper yet.

Whiteboard

No one has generated a whiteboard explanation for this paper yet.

Open Problems

We haven't generated a list of open problems mentioned in this paper yet.

Collections

Sign up for free to add this paper to one or more collections.

Tweets

Sign up for free to view the 1 tweet with 0 likes about this paper.