Improving Quantized Model Performance in Qualitative Analysis with Multi-Pass Prompt Verification
Researchers have developed a multi-pass prompt verification method to improve the accuracy of quantized Large Language Models (LLMs) in qualitative analysis. The study focused on LLaMA-3.1 (8B) models quantized to various bit levels (8-bit, 4-bit, 3-bit, and 2-bit), finding that lower bit levels often lead to increased hallucinations and instability. The proposed method guides the model through controlled steps to reduce unreliable content, significantly enhancing the performance of 4-bit models and improving even the heavily compressed 3-bit and 2-bit models. AI
IMPACT Enhances the usability of resource-efficient LLMs for qualitative research, potentially lowering costs and increasing accessibility.