HILL: A Hallucination Identifier for Large Language Models

Document Title

HILL: A Hallucination Identifier for Large Language Models

Document Information

• Subject Area: Human-Computer Interaction and Artificial Intelligence (specifically addressing the "hallucination" phenomenon in large language models)
• Keywords: ChatGPT, Large Language Models, Artificial Hallucination, User-Centered Design, Hallucination Detection

Research Background and Problem

• Identified Problems or Challenges:
  • Large Language Models (LLMs, such as ChatGPT) are prone to hallucinations—unreliable, inaccurate, or nonsensical text outputs. These hallucinations can lead to user misunderstandings or poor decision-making, especially when users overly rely on these outputs. Furthermore, existing solutions primarily focus on reducing the model's own errors, without adequately addressing how user interface design can help users detect and respond to hallucinations.
• Significance of the Problem:
  • As LLMs are increasingly adopted in daily life, their impact could extend across various domains, such as political opinions and educational cognition. Addressing users' over-reliance on LLMs is crucial, and designing tools and interfaces that help users identify hallucinations is key to this effort.
• Research Motivation and Related Work:
  • Some studies have attempted to reduce LLM errors through technical means, but hallucinations cannot be entirely avoided in the future. The necessity of this research lies in improving user interaction design to enable users to recognize hallucinations and take appropriate actions.

Solution

• Proposed Method or Solution:
  • HILL (Hallucination Identifier for Large Language Models): A novel human-computer interaction tool that intuitively helps users identify hallucinations in LLM outputs through a user-friendly interface. HILL is developed based on the ChatGPT API and incorporates various design features (e.g., confidence scores, source links, and highlighted hallucinated content).
• Innovative Aspects:
  • HILL's design is based on user-centered principles, combining interactive prototypes and user feedback to visualize hallucinations in the interface, thereby enhancing user comprehension and experience.
  • It not only focuses on correcting the model's errors but also emphasizes user behavior interventions, prioritizing hallucination recognition over complete reliance on model performance.
• Implementation Steps and Key Technologies:
  • Prototype Design and Evaluation: Development of three interactive prototypes, evaluated through Wizard of Oz experiments to prioritize user-desired design features.
  • Feature Optimization and Integration: Integration of key features (e.g., confidence scores, hallucination markers, source links) into a complete application based on user feedback.
  • Technical Implementation: Frontend developed using the Vue.js framework, backend based on Express.js for communication with the ChatGPT API, supporting multiple API requests to verify response accuracy.
  • Confidence Score Calculation: Overall confidence is calculated using a weighted method that combines self-assessment scores, source quality, and hallucination markers.

Research Outcomes

• Specific Achievements:
  • Improved User Experience: Compared to the traditional ChatGPT interface, HILL significantly enhances users' ability to detect hallucinations and better draws their attention to potential misinformation.
  • Performance Evaluation: Testing on the Stanford Question Answering Dataset (SQuAD 2.0) shows that HILL effectively detects hallucinations (achieving a recall rate of 60%-66.7% for hallucinations in incorrect answers).
  • Balancing User Trust: Users demonstrated high trust in HILL but also highlighted potential risks of over-reliance.
• Comparative Advantages Over Existing Solutions:
  • HILL supplements existing purely technical approaches by involving users directly in hallucination detection through clear interface and interaction design.
• Limitations and Future Directions:
  • Limitations:
    • Small sample size; user experience evaluation was primarily conducted through video demonstrations rather than in natural interaction environments.
    • Some design features (e.g., confidence thresholds) were not fully implemented, requiring further exploration for technical realization.
    • Current reliance on the ChatGPT API for self-assessment introduces communication complexity and high costs.
  • Future Directions:
    • Expand the scale of user evaluations to study the tool's long-term impact on user behavior.
    • Explore more optimized interface designs and incorporate additional customizable features (e.g., dynamic confidence thresholds).
    • Leverage external verification mechanisms (e.g., search engines) to further improve hallucination detection accuracy and optimize source quality analysis.

Summary: This study developed a user-centered tool, HILL, to help users identify hallucinations in large language models. Through interface optimization and feature integration, HILL demonstrates significant effectiveness in enhancing users' scrutiny and accuracy when interacting with LLM outputs, while providing a guiding framework for future user interaction design in artificial intelligence.