FiNE: Fine-grained Neuron-level Model Editing for Reliable and Safe LLMs.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Large Language Models (LLMs) require enhanced reliability and safety measures.
• Existing editing methods like ROME may overlook internal network structures, causing brittle updates.
• There is a need for more robust and fine-grained editing techniques for LLMs.

Purpose of the Study:

• To propose FiNE (Fine-grained Neuron-level Editing), a unified framework for LLM editing.
• To address limitations of prior methods by focusing on internal Feed-Forward Network (FFN) structures.
• To enhance both factual accuracy and safety of LLM outputs.

Main Methods:

• FiNE employs gradient-free localization and fine-grained parameter adjustment within FFNs.
• A theoretically grounded contribution score identifies concept-relevant neurons, acting as an efficient alternative to gradient-based methods.
• A composite loss function balances editing accuracy, model consistency, and output diversity.

Main Results:

• FiNE demonstrates superior precision, robustness, and efficiency in knowledge and safety editing tasks.
• Experiments on KnowEdit and SafeEdit benchmarks show minimal disruption to the LLM's general behavior.
• FiNE achieves state-of-the-art results on both factual correction and safety enhancement.

Conclusions:

• Fine-grained neuron editing is an effective and scalable approach for improving LLMs.
• FiNE offers a unified framework for building more reliable and safer Large Language Models.
• The proposed method advances the field of LLM interpretability and controllability.