PaperlayerView exact revision on arXiv

Source-linked AI summary

HealMe: Harnessing Cognitive Reframing in Large Language Models for Psychotherapy

Mengxi Xiao, Qianqian Xie, Ziyan Kuang, Zhicheng Liu, Kailai Yang, Min Peng, Weiguang Han, Jimin Huang

arXiv:2403.05574v3cs.HCcs.AIcs.CL

TL;DR

Prior cognitive-reframing LLMs often converted negative emotions into positive expressions without encouraging clients’ discovery of alternative perspectives. HealMe uses empathetic, guided dialogue to help clients reframe their thoughts. The paper reports stronger empathy, guidance, and coherence than existing models, with some real-person clients showing lower negative emotional scores.

  • Problem

    Prior LLM approaches often treated cognitive reframing as rewriting negative emotions, which may not promote clients’ self-discovery of alternative perspectives.

    Source: §1 · p. 1

  • Method

    HealMe uses empathetic dialogue to guide clients through separating situations from thoughts, brainstorming alternatives, and considering actionable suggestions.

    Source: §1 · p. 2

  • Results

    HealMe is reported to surpass other models in empathy, guidance, and coherence, while some real-person clients showed decreases in negative emotional attributes.

    Source: §1 · p. 3, §8 · p. 9

  • Takeaways & Limitations

    The findings indicate HealMe’s potential to support psychotherapy through AI-enhanced methods and client-directed cognitive reframing.

    Source: §8 · pp. 8–9

    Dialogue is limited to three rounds, and the structured stepwise prompts constrain flexibility when clients face multifaceted issues.

    Source: §9 · p. 9

Abstract

from arXiv · show

Large Language Models (LLMs) can play a vital role in psychotherapy by adeptly handling the crucial task of cognitive reframing and overcoming challenges such as shame, distrust, therapist skill variability, and resource scarcity. Previous LLMs in cognitive reframing mainly converted negative emotions to positive ones, but these approaches have limited efficacy, often not promoting clients' self-discovery of alternative perspectives. In this paper, we unveil the Helping and Empowering through Adaptive Language in Mental Enhancement (HealMe) model. This novel cognitive reframing therapy method effectively addresses deep-rooted negative thoughts and fosters rational, balanced perspectives. Diverging from traditional LLM methods, HealMe employs empathetic dialogue based on psychotherapeutic frameworks. It systematically guides clients through distinguishing circumstances from feelings, brainstorming alternative viewpoints, and developing empathetic, actionable suggestions. Moreover, we adopt the first comprehensive and expertly crafted psychological evaluation metrics, specifically designed to rigorously assess the performance of cognitive reframing, in both AI-simulated dialogues and real-world therapeutic conversations. Experimental results show that our model outperforms others in terms of empathy, guidance, and logical coherence, demonstrating its effectiveness and potential positive impact on psychotherapy.

1 Introduction

HealMe reframes negative thoughts through empathetic dialogue that guides clients toward self-discovery rather than simply rewriting negative emotions. The paper reports stronger empathy, guidance, and logical coherence than other models in simulated dialogue, and decreases in negative emotional scores for some real-person clients.

1 Introduction

  • LLMs may help address shame, distrust, limited psychotherapy resources, and variability in therapist skill.Cognitive reframing is described as a key part of CBT that helps people detach from thoughts and situations.Source: §1 · p. 1
  • HealMe guides clients to distinguish situations from thoughts, brainstorm alternative perspectives, and receive empathetic suggestions that recognize effort and encourage action.Its dialogue-based approach contrasts with sentence rewriting, which may not prompt clients to discover alternatives themselves and can feel imposed.Source: §1 · pp. 1–2
  • HealMe outperforms other models in empathy, guidance, and logical coherence in AI-to-AI dialogues; some real-person clients’ negative scores fell from 5/5 to 1/5.The evaluation uses a three-dimensional metric for simulated conversations and professional psychological metrics for real-person scenarios.Source: §1 · p. 3
  • The paper presents HealMe as a way to sustain empathy and guidance in LLM psychotherapy, with evaluation metrics for public and non-public therapy dialogues.The authors compare their approach with other LLMs in simulated conversations and human interactions.Source: §1 · p. 3

2 Problem Definition and Goals

The paper describes cognitive reframing dialogues as guiding clients out of cognitive traps, while evaluating simulated and real-client therapy differently.

2 Problem Definition and Goals

  • For AI-simulated clients, therapist responses are assessed for empathy, logical consistency, and guidance; for real clients, effectiveness is assessed through changes in emotional attributes before and after therapy.Source: §2 · p. 3

3 Dataset Construction

HealMe’s dialogue dataset expands cognitive-reframing examples into a three-step therapy process, with simulated client and therapist turns reviewed and used for training.

3 Dataset Construction

  • The dataset starts from 1,000 manually selected, well-composed pairs of thinking traps and client thoughts from an existing cognitive-reframing dataset.The underlying dataset represents common thinking traps and thoughts generated by crowd workers with psychology backgrounds.Source: §3 · p. 3

3.1 Step 1: Separating Emotions from Facts

  • In step one, clients express their thoughts and confusion, then the therapist guides them to separate situations from thoughts.Source: §3.1 · pp. 3–4

3.2 Step 2: Brainstorming

  • In step two, the therapist guides brainstorming under a given situation, aiming for different viewpoints rather than a perfect reframe.The client is prompted to brainstorm, but some examples deliberately remain negative to reflect difficulty finding neutral or positive possibilities while distressed.Source: §3.2 · p. 4, §3.3 · p. 4

3.3 Step 3: Empathetic Response

  • In step three, the therapist acknowledges the client’s reframing effort and responds with situation-specific empathy that addresses the thinking trap.The generated three-step dialogues and therapist prompts are used for training.Source: §3.3 · p. 4

4 Dataset Evaluation

The paper evaluates training dialogues with client-role checks and therapist-quality ratings, reporting high empathy and logic in the training set and guidance around 2 as a balance.

4.1 Evaluation of the AI Client

  • AI-client responses are checked for clear expression of situation and emotion, staying in role, and following therapist instructions across the dialogue.Responses are generated and revised until they meet all three criteria; inability to brainstorm while sad can still count as following instructions.Source: §4.1 · p. 4

4.2 Evaluation of the AI Therapist

  • Therapist replies are rated for empathy, logical coherence, and guidance, with each metric scored from 0 to 3.Domain experts manually review 70 random dialogues, then use those examples and metrics to guide GPT-4 scoring of the training set.Source: §4.2 · pp. 4–5
  • The training set is judged high quality, with high empathy and strong logic; guidance around 2 is described as a balance that avoids overly directive, potentially preachy advice.Source: §4.2 · p. 5, Table 2 · p. 5

5 Training

HealMe is trained by fine-tuning LLaMA2-7b-chat on the training data, with an initial 100 multi-round conversations set aside for validation.

5 Training

  • The initial 100 multi-round training conversations are partitioned to form the validation set.Source: §5 · p. 5
  • HealMe is trained for three epochs on LLaMA2-7b-chat, and the best-performing model is selected using validation results.The reported training time is 2h 12m 44s.Source: §5 · p. 5
  • Training uses AdamW with a maximum learning rate of 3e-4 and a warmup ratio of 1%.Source: §5 · p. 5
  • Training runs on four Nvidia GeForce RTX 3090 GPUs, each with 24GB of memory.Source: §5 · p. 5

6 Experiments

The experiments evaluate HealMe against ChatGLM3-6b and LLaMA2-7b-chat in AI-client dialogues and a small real-client test. HealMe performs best across the evaluated AI-dialogue categories, while real-client cases show reported emotional improvements in some instances.

6.3 Analysis of Experimental Results

  • HealMe outperforms ChatGLM3-6b and LLaMA2-7b-chat across all evaluated categories in AI-client conversations.The evaluation covers empathy, logical coherence, guidance, and an overall score; two psychologists anonymously score dialogues in random order. The test uses 300 cases and ChatGPT-simulated clients over three rounds.Source: §6.2 · pp. 5–6, §6.3 · p. 6, Table 4 · p. 6
  • HealMe’s higher empathy, coherence, and guidance scores are associated in the authors’ discussion with emotional understanding, structured dialogue, and constructive feedback.The authors describe actionable advice as supporting problem-solving and self-reflection.Source: §6.3 · p. 6

6.4 Case Study

  • For a challenging case where the client is too sad to brainstorm alternatives, HealMe receives the highest score and a full empathy score.The authors also describe stronger interactivity in the Figure 2 dialogue.Source: §6.4 · p. 6, Table 5 · p. 6
  • In reviewed low-scoring cases, the lowest observed score was 1 on a 0–3 scale, and the authors report no identified potential negative impacts.They say even poor-performing responses provided basic, general comfort.Source: §6.4 · p. 6

6.5 A Supplementary Test with Real Person Clients

  • The supplementary real-client test assigns six volunteers to three models and limits client interactions to three rounds.Two additional clients form a control group; the test concerns short-term negative experiences lasting more than a day but less than a week.Source: §6.5 · p. 7
  • In some real-client cases, HealMe is reported to reduce negative emotions and sensitivity while increasing determination.Examples include clients reconsidering effort and interpreting others’ actions as not necessarily caused by their own mistakes.Source: §6.5 · p. 8

7 Related Work

Related work examines LLM empathy and cognitive reframing, while HealMe is positioned as a fine-tuned therapist model that implements a broader reframing process.

7 Related Work

  • A prior comparison found evaluators preferred ChatGPT responses to physicians’ social-media responses expressing negative emotions 78.6% of the time.The passage notes that empathy alone may not meet psychotherapy’s need for guidance, and positions fine-tuned LLaMA2-7b-chat as supporting consistent therapist-role, empathy, and guidance.Source: §7 · p. 8
  • LLM counseling platforms and psychiatry studies also explore mental-health applications, but the passage distinguishes HealMe by its fine-tuning for therapist-role consistency and guidance.Source: §7 · p. 8
  • Prior LLM cognitive-reframing work primarily rewrites negative emotions, whereas HealMe implements cognitive reframing as a broader psychotherapy process (Ziems et al., 2022; Maddela et al., 2023; Sharma et al., 2023b).Source: §7 · p. 8

8 Conclusion

The conclusion presents HealMe as an empathetic, psychotherapeutically grounded approach to cognitive reframing that promotes balanced perspectives and client self-discovery.

8 Conclusion

  • HealMe applies cognitive reframing through empathetic dialogue grounded in psychotherapeutic principles to address deep-rooted negative thoughts and promote balanced perspectives.Source: §8 · p. 8
  • The approach emphasizes clients’ self-discovery and rational thought processes rather than only converting negative emotions.Source: §8 · p. 8
  • The paper reports that its comprehensive psychological evaluation metrics confirm HealMe’s superiority over existing models in empathy, guidance, and coherence.The authors describe this as signifying potential for AI-enhanced psychotherapy.Source: §8 · p. 9

9 Limitations

HealMe’s three-round, structured approach can leave some concerns unresolved when clients bring multifaceted issues. The authors identify limited flexibility as a related constraint.

9 Limitations

  • Three dialogue rounds can leave concerns unresolved when clients face multifaceted issues.The authors say the step-by-step approach enhances specificity but that structured prompts restrict flexibility.Source: §9 · p. 9
  • HealMe’s structured prompts restrict the flexibility of its step-by-step guidance.The authors plan to explore flexible dialogue rounds and a broader range of psychotherapeutic strategies.Source: §9 · p. 9

10 Ethical Considerations

The paper describes privacy protections for real-person conversations and the expertise of its evaluators. Its AI-to-AI assessment scores empathy, logical coherence, guidance, and an overall rating.

10 Ethical Considerations

  • Real-person participants could withdraw at any time, and their data were deleted if they did; conversations were confidential and not recorded after exit.No participants withdrew during the experiments.Source: §10 · p. 9
  • The evaluation team comprised two co-author experts, including a mental health professional with 11 years of counseling-research experience.The main evaluator established the guidelines and participated in scoring; the second evaluator also engaged in the evaluation.Source: §10 · p. 9

A Evaluation Metrics in AI-to-AI Conversations

B.1 Case Study in Guidance

The case studies contrast HealMe’s confirmation, targeted brainstorming, and responses to client pain with baseline-model replies. The appendix describes the real-person evaluation setup, PANAS procedure and findings, and client feedback on baseline models.

B.1 Case Study in Guidance

  • Across rounds, HealMe alone checks its analysis, offers targeted brainstorming examples, and responds to pain with validation and sadness-facing suggestions.The other models affirm the client without checking, do not guide brainstorming despite the prompt, and focus on distraction or contacting others in the final round.Source: §B.1 · p. 13

B.2 A Low-scoring Case of HealMe

  • In a low-scoring case, HealMe’s guidance opened the client’s thoughts but still directed brainstorming toward attractiveness at work; the client ultimately identified their attractiveness.Evaluators assigned the case scores of 1211 and 1111 for empathy, logical coherence, guidance, and overall performance, respectively, as printed in the passage.Source: §B.2 · p. 15

C Evaluation Details in AI-to-Human Conversations

  • Because of resource constraints, the real-person experiments were small-scale and collected PANAS changes and client feedback while protecting client privacy.The appendix directs readers to its pre-test, procedures, and specific-score sections for further details.Source: C Evaluation Details in AI-to-Human… · p. 16

C.1 Pre-test Evaluation

  • The three client groups showed no significant pre-test difference (p-value 0.164; h0 = 0.065), supporting random assignment of therapy models.Source: §C.1 · p. 16

C.2 Evaluation Metrics in Real Conversations

  • The real-conversation measure was a 20-question PANAS questionnaire completed before clients interacted with an AI psychotherapy model.A domain-expert co-author introduced PANAS and guided questionnaire completion before leaving the room.Source: §C.2 · p. 16

C.2.1 Testing Procedure

  • The testing procedure guided clients through rating emotions from the past week on a 1–5 scale, with support and confidentiality assurances.Clients were encouraged to answer honestly and complete the questionnaire at their own pace.Source: §C.2.1 · p. 16

C.3 Clients PANAS Score changes

  • The authors report a correlation between model use and negative-emotion fluctuations and interpret the observations as evidence of alleviation.They describe pre- and post-intervention PANAS scores for eight clients and define fluctuation as the absolute negative-emotion change divided by the pre-experiment negative-emotion total.Source: §C.3 · pp. 16–17

C.4 Client’s feedback on baseline models

  • Client feedback described LLaMA2-7b-chat as cold and insufficiently empathetic, while ChatGLM3-6b’s advice satisfied two clients but felt insufficiently guided to one.One client reported reduced fear and a more objective view after following LLaMA2-7b-chat’s advice.Source: §C.4 · p. 17

Related papers

Loading 2403.05574v3…