This research introduces Post-quantization Integral (PQI), a method to identify sensitive weights in Large Language Models, enabling more efficient model compression without sacrificing performance.

• Proposes PQI, a metric that measures weight sensitivity based on loss function impact
• Introduces ReQuant, a preprocessing technique that preserves sensitive weights during quantization
• Demonstrates significant improvements over existing quantization methods across multiple LLMs
• Reduces computational costs while maintaining model accuracy and capabilities

Why it matters: This engineering advancement makes deploying large AI models more affordable and accessible by reducing memory requirements and computational costs, while maintaining model quality.

Identifying Sensitive Weights via Post-quantization Integral