CatAlyst: Domain-Extensible Intervention for Preventing Task Procrastination Using Large Generative Models

Title of the Paper

CatAlyst: Domain-Extensible Intervention for Preventing Task Procrastination Using Large Generative Models

Paper Information

• Research Area: Technology and Digital Well-being, AI-Assisted Task Management, Human-Computer Interaction
• Keywords: Large Generative Models, Behavior Change, Task Engagement, Procrastination, Human-AI Collaboration

Research Background and Problem Statement

• Problem or Challenge:

  • Modern knowledge workers often lose focus while performing high cognitive load tasks (e.g., writing, programming, editing slides), leading to procrastination and reduced productivity.
  • Traditional AI-assisted tools focus on directly completing tasks, but their performance is often limited by model accuracy, and low-quality outputs can undermine users' trust in the system.
  • Customizing and adapting generative models for each specific task is prohibitively expensive and difficult to achieve.

• Significance:

  • Procrastination not only decreases work efficiency but also leads to stress, reduced self-efficacy, and negative impacts on digital well-being.

• Research Motivation and Related Work:

  • Many current studies support human-AI collaboration, but most focus on precise content generation and fail to explore the behavioral inducement effects of generated content.
  • HCI literature has proposed various intervention techniques, such as motivational messages and context-aware feedback, but their effectiveness is limited in high cognitive load tasks.

Solution

• Method or Solution:

  • Propose a novel AI system design, CatAlyst, which uses large generative models to generate content for resuming interrupted tasks, inducing behavioral changes to enhance user task engagement.
  • CatAlyst is domain-extensible, eliminating the need for task-specific model customization, thus improving deployment cost-effectiveness.

• Innovative Features:

  • Focus on the application of generative models for inducing behavioral changes rather than directly replacing parts of the task.
  • Generate personalized content for task resumption as an intervention method, attracting users to re-engage with tasks while reducing cognitive load during task recovery.

• Implementation Steps and Key Technologies:

  • Automatically detect moments of task interruption by identifying loss of focus through interaction logs.
  • Use large generative models (e.g., GPT-3) to generate task resumption content and present it to users as an intervention.
  • Deliver interventions via notifications, leveraging generated content to lower psychological barriers to task recovery.
  • Support multi-domain tasks, such as writing and slide editing, with simple model input adjustments for applicability.

Research Findings

• Specific Outcomes:

  • CatAlyst effectively attracts user attention through content-based interventions and reduces cognitive load during task recovery.
  • Users recover tasks faster and demonstrate increased task focus after using the system.

• Comparison with Existing Solutions:

  • In comparative studies with traditional intervention methods (e.g., motivational messages), participants resumed tasks more quickly and reported better subjective experiences after CatAlyst interventions.
  • Long-term experiments show that the novelty of generated content consistently captures user interest, whereas traditional inducement methods often fail due to repetition.

• Experimental or Evaluation Results:

  • In writing tasks, users resumed attention significantly faster with CatAlyst interventions compared to traditional methods (18.8 seconds vs. 135.0 seconds).
  • In slide editing tasks, generated content helped users reduce cognitive load and task completion time (1356.5 seconds vs. 2794.5 seconds without intervention).
  • Users perceived the system as providing roles such as reminders, ideation support, and collaboration partners.

• Limitations and Future Directions:

  • The logic for detecting task interruption moments is relatively simple (based on time thresholds), leaving room for improvement in accuracy.
  • The quality of generated content in certain domains is inconsistent, affecting user experience.
  • Long-term behavioral impact studies require more systematic approaches.
  • Ethical concerns and potential biases in generative models need to be evaluated before system deployment.
  • Exploring integration with other intervention methods (e.g., HabitLab's rotating intervention model) could maintain novelty and coherence in interventions.