MindShift: Leveraging Large Language Models for Mental-States-Based Problematic Smartphone Use Intervention

Title of the Paper

MindShif: Leveraging Large Language Models for Mental-States-Based Problematic Smartphone Use Intervention

Paper Information

• Field of Study: Human-Computer Interaction and Behavior Change Technology
• Keywords: Problematic smartphone use, persuasive strategies, large language models, psychological models, behavioral intervention

Research Background and Problem

• Identified Issues or Challenges:

  • Excessive smartphone use negatively impacts physical and mental health, leading to reduced efficiency, decreased sleep, and mental health problems.
  • Current intervention techniques (e.g., self-monitoring, reminders, interaction blocking, and locking) have the following shortcomings:
    • Lack of dynamism, unable to adjust content based on users' real-time context and psychological state.
    • Overly templated reminder content reduces user acceptance.
    • Insufficient focus on users' intrinsic psychological states (e.g., stress, boredom, inertia).

• Research Motivation:

  • Existing smartphone use interventions primarily focus on external physical contexts, neglecting users' instantaneous psychological states.
  • Behavioral interventions based on psychological states may be more effective and require dynamic adjustment of intervention content.

• Related Work:

  • Review of existing intervention methods, including external interventions (reminders, interaction blocking) and internal design interventions (revised interface design).
  • Research on the relationship between psychological states (e.g., stress, boredom) and user behavior.

Solution

• Proposed Method or Solution:

  • MindShift System: Utilizing large language models (LLMs) to generate dynamic, personalized persuasive content, integrating users' real-time usage data, physical context, psychological state, goals, and habits.
  • Psychological-State-Driven Persuasive Strategies: Based on dual-system theory and growth needs theory, four persuasive strategies are summarized:
    1. Understanding: Expressing empathy for users' emotions.
    2. Comforting: Soothing users' emotional fluctuations.
    3. Evoking: Reminding users of personal goals.
    4. Scaffolding Habits: Helping users replace habitual behaviors with positive habits.

• Innovations:

  • Considering users' real-time psychological states (e.g., boredom, stress, inertia) and activity engagement (whether engaged in tasks).
  • Using LLMs to generate dynamic, context-specific persuasive content.
  • Developing a complete user interaction process: initializing custom goals and habits, real-time reporting of user intentions and psychological states, and displaying personalized content.

• Implementation Steps and Key Technologies:

  1. Designing detailed prompts for LLMs, providing user context and strategy descriptions.
  2. Developing a mechanism combining automation and user self-reporting to collect psychological state data.
  3. Integrating LLMs to generate real-time intervention messages and dynamically display them.
  4. Evaluating the system's effectiveness through experiments.

Research Outcomes

• Specific Results:

  • In a five-week field experiment, MindShift significantly increased user acceptance of intervention content (by 4.7%-22.5%) and reduced smartphone usage duration (by 7.4%-9.8%).
  • After using MindShift, smartphone addiction scores decreased by 34.7%, and self-efficacy scores increased by 10.7%.

• Advantages Compared to Existing Solutions:

  • MindShift combines users' psychological states and goals, enhancing the dynamism and personalization of interventions.
  • Persuasive strategies and LLM-generated content make interventions more appealing.

• Experimental or Evaluation Results:

  • Among three methods (MindShift, MindShift-Simple, and Baseline), MindShift performed best in acceptance rate (35.2%) and user satisfaction (6.8% positive feedback rate).
  • MindShift significantly reduced habitual smartphone use behaviors related to psychological states (e.g., boredom) and activity engagement levels.

• Limitations and Future Directions:

  • Limitations:
    • Participants were primarily young individuals, limiting sample diversity.
    • The experiment duration was relatively short (five weeks), preventing exploration of long-term intervention effects.
    • Stability and privacy concerns of LLM-generated content require further optimization.
  • Future Directions:
    • Enhancing the diversity and personalization of intervention content, developing adaptive intervention systems based on user feedback.
    • Exploring LLM-driven dynamic intervention technologies in other behavior change domains (e.g., smoking cessation, dietary habit formation).
    • Researching automated alternatives to user self-reported psychological state data to reduce user burden while ensuring personalization and dynamism of interventions.