Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

PanicToCalm: A Proactive Counseling Agent for Panic Attacks

Jihyun Lee, Yejin Min, San Kim, Yejin Jeon, SungJun Yang, Hyounghun Kim, Gary Geunbae Lee

TL;DR

The paper presents Pace, the first dataset grounded in Psychological First Aid (PFA) for acute panic scenarios, enabling training of PACER, a crisis-focused counseling model. PACER is trained with supervised fine-tuning and simulated preference alignment (DPO) to deliver rapid stabilization and directive support, evaluated via PanicEval and PANAS across simulated and human assessments. Results show PACER achieves strong stabilization and desirable client affect improvements, often preferred to general, CBT-based, and GPT-4o baselines, while revealing a tradeoff between empathy and directive efficacy. The work provides an end-to-end pipeline ( Pace, PACER, PanicEval) for developing AI systems capable of first-aid-oriented psychological support in high-distress contexts, while noting real-time deployment, privacy, and consent considerations as future work.

Abstract

Panic attacks are acute episodes of fear and distress, in which timely, appropriate intervention can significantly help individuals regain stability. However, suitable datasets for training such models remain scarce due to ethical and logistical issues. To address this, we introduce PACE, which is a dataset that includes high-distress episodes constructed from first-person narratives, and structured around the principles of Psychological First Aid (PFA). Using this data, we train PACER, a counseling model designed to provide both empathetic and directive support, which is optimized through supervised learning and simulated preference alignment. To assess its effectiveness, we propose PanicEval, a multi-dimensional framework covering general counseling quality and crisis-specific strategies. Experimental results show that PACER outperforms strong baselines in both counselor-side metrics and client affect improvement. Human evaluations further confirm its practical value, with PACER consistently preferred over general, CBT-based, and GPT-4-powered models in panic scenarios (Code is available at https://github.com/JihyunLee1/PanicToCalm ).

PanicToCalm: A Proactive Counseling Agent for Panic Attacks

TL;DR

The paper presents Pace, the first dataset grounded in Psychological First Aid (PFA) for acute panic scenarios, enabling training of PACER, a crisis-focused counseling model. PACER is trained with supervised fine-tuning and simulated preference alignment (DPO) to deliver rapid stabilization and directive support, evaluated via PanicEval and PANAS across simulated and human assessments. Results show PACER achieves strong stabilization and desirable client affect improvements, often preferred to general, CBT-based, and GPT-4o baselines, while revealing a tradeoff between empathy and directive efficacy. The work provides an end-to-end pipeline ( Pace, PACER, PanicEval) for developing AI systems capable of first-aid-oriented psychological support in high-distress contexts, while noting real-time deployment, privacy, and consent considerations as future work.

Abstract

Panic attacks are acute episodes of fear and distress, in which timely, appropriate intervention can significantly help individuals regain stability. However, suitable datasets for training such models remain scarce due to ethical and logistical issues. To address this, we introduce PACE, which is a dataset that includes high-distress episodes constructed from first-person narratives, and structured around the principles of Psychological First Aid (PFA). Using this data, we train PACER, a counseling model designed to provide both empathetic and directive support, which is optimized through supervised learning and simulated preference alignment. To assess its effectiveness, we propose PanicEval, a multi-dimensional framework covering general counseling quality and crisis-specific strategies. Experimental results show that PACER outperforms strong baselines in both counselor-side metrics and client affect improvement. Human evaluations further confirm its practical value, with PACER consistently preferred over general, CBT-based, and GPT-4-powered models in panic scenarios (Code is available at https://github.com/JihyunLee1/PanicToCalm ).
Paper Structure (76 sections, 1 equation, 7 figures, 20 tables)

This paper contains 76 sections, 1 equation, 7 figures, 20 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data Challenges in Mental Health
  4. Synthetic Preference Pairs for Value Alignment
  5. Psychological First Aid
  6. Pace: Panic Episode Dataset with PFA Counseling
  7. Collecting Client Panic Profiles
  8. Defining Counselor Goals
  9. Structured Dialogue Generation with PFA
  10. Dialogue Quality Filtering
  11. Expert Review of Pace
  12. Comparison Pace with Existing Counseling Datasets
  13. Pacer: Panic Counseling Model with PFA
  14. Experiments
  15. Training Details
  16. ...and 61 more sections

Figures (7)

  • Figure 1: Conversation between a client and different counseling models. Pacer delivers both appropriate empathy and clear, directive guidance to support immediate stabilization.
  • Figure 2: An overview of the automatic pipeline for generating the Pace dataset.
  • Figure 3: Illustration of the preference alignment process using DPO.
  • Figure 4: Client's change in emotions before and after counseling, based on PANAS subscales.
  • Figure 5: Head-to-head comparisons of Pacer against PacerSFT, Camel, and GPT-4o using Gemini-based client simulator and evaluator.
  • ...and 2 more figures