<ul><li>Supervised fine-tuning of Large Language Models (LLMs) on high-quality datasets can improve reasoning capabilities.</li><li>Full fine-tuning (Full FT) is powerful but computationally expensive and prone to overfitting, especially with limited data.</li><li>Sparse fine-tuning, focusing on updating a small subset of important model parameters, strikes a balance between efficiency and effectiveness.</li><li>A new method called Low-rank Informed Sparse Fine-Tuning (LIFT) identifies Principal Weights, crucial for reasoning, through rank reduction, outperforming Full FT on reasoning tasks.</li></ul>

LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning

Discover more