Knowledge Editing and Model Editing Research

What is Knowledge Editing in Large Language Models?

Knowledge Editing is a technique to modify and correct factual knowledge encoded in Large Language Models (LLMs) without retraining from scratch. It allows for precise, efficient modifications to fix hallucinations, update outdated information, or inject new knowledge into pre-trained models.

What is Model Editing?

Model Editing is an emerging research area that enables precise and efficient modifications to Large Language Models while preserving their overall capabilities. It encompasses knowledge editing, behavior editing, and other techniques to update model parameters or behavior without full retraining.

Can Knowledge Editing correct hallucinations in LLMs?

Yes, Knowledge Editing can correct hallucinations in LLMs. Our HalluEditBench benchmark demonstrates that various knowledge editing methods can effectively correct real-world hallucinations across multiple dimensions including Efficacy, Generalization, Portability, Locality, and Robustness. However, the effectiveness varies by method and context.

What is HalluEditBench?

HalluEditBench is a comprehensive benchmark for evaluating knowledge editing methods in correcting real-world hallucinations in Large Language Models. It assesses performance across five dimensions: Efficacy, Generalization, Portability, Locality, and Robustness, with over 6,000 hallucinations across 9 domains and 26 topics.

What is Behavior Editing?

Behavior Editing is a novel paradigm that frames ethical behavior steering of AI agents as a model editing task. Using the BehaviorBench benchmark grounded in psychological moral theories, Behavior Editing can precisely steer both benevolent and harmful behaviors in LLM-based agents.

What is Personalization Editing?

Personalization Editing is a framework that applies localized edits guided by clustered preference representations to align LLMs with individual user preferences. It enables precise preference-aligned updates while preserving overall model capabilities, offering higher efficiency than fine-tuning and better persistence than prompting-based methods.

What are Editing Attacks?

Editing Attacks refer to the malicious use of knowledge editing techniques to inject misinformation or bias into Large Language Models. Research shows that editing attacks can successfully inject harmful content with up to 90% success rate, highlighting new safety threats in the era of editable AI systems.