<ul><li>ULTRAEDIT is a novel solution for model editing in large language models (LLMs) that is training-, subject-, and memory-free, making it well-suited for real-world lifelong model editing.</li><li>ULTRAEDIT performs editing using lightweight linear algebra operations to compute parameter shifts efficiently, achieving editing speeds over 7 times faster than the previous state-of-the-art method.</li><li>It employs a lifelong normalization strategy to adapt to distributional shifts and maintain consistency over time, enabling it to edit a 7B LLM on a 24GB consumer-grade GPU with minimal VRAM consumption.</li><li>Experiments on diverse datasets and models demonstrate that ULTRAEDIT consistently outperforms other methods in model editing scenarios, supporting up to 1M edits while maintaining high accuracy.</li></ul>

UltraEdit: Training-, Subject-, and Memory-Free Lifelong Editing in Large Language Models

Discover more