<ul data-eligibleForWebStory="true"><li>Researchers from multiple universities introduce Drag-and-Drop (DnD) LLMs to customize large language models quickly.</li><li>DnD generates task-specific LoRA adapters from prompts, offering faster and more accurate results than traditional methods.</li><li>The new approach involves combining a frozen text encoder with a hyper-convolutional decoder for efficient adapter weight generation.</li><li>DnD collapses the conventional 'data→gradients→weights' loop into a single forward step, challenging the necessity of gradient descent for model specialization.</li><li>Compared to traditional fine-tuning methods like LoRA, DnD provides task-specific parameters up to 12,000 times faster and achieves up to 30% performance gains.</li><li>DnD significantly enhances accuracy on various datasets including ARC-e, BoolQ, HumanEval, GSM8K, and Math-Vision.</li><li>The method generalizes well across different domains and model sizes, demonstrating improved accuracy even on datasets it wasn't specifically trained on.</li></ul>

New Method Customises LLMs in Seconds, Beats Tuning: Research

Discover more