KEditVis: A Visual Analytics System for Knowledge Editing of Large Language Models

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction

Background:

• Large Language Models (LLMs) excel at factual question answering but can produce inaccuracies.
• Knowledge editing techniques aim to correct factual errors in LLMs.
• Current knowledge editing methods lack transparency in layer selection and rely on insufficient guidance indicators.

Purpose of the Study:

• To present KEditVis, a visual analytics system for understanding and improving knowledge editing in LLMs.
• To enable users to identify optimal model layers for editing and diagnose ineffective edits.
• To facilitate the discovery of insights for advancing knowledge editing algorithms.

Main Methods:

• Development of KEditVis, a novel visual analytics system.
• Integration of interactive visualizations to explore knowledge editing processes.
• Evaluation through usage scenarios, expert interviews, and a user study.

Main Results:

• KEditVis provides users with deeper understanding of knowledge editing.
• The system aids in selecting appropriate editing targets and performing more effective edits.
• Evaluation confirms the system's effectiveness and usability.

Conclusions:

• KEditVis enhances transparency and effectiveness in LLM knowledge editing.
• Interactive visualization is crucial for optimizing knowledge editing strategies.
• The system supports targeted factual correction and informs future algorithm development.