Reproduce “Model Editing Harms General Abilities of Large Language Models: Regularization to the Rescue” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jia-Chen Gu, Hao-Xiang Xu, Jun-Yu Ma, Pan Lu, Zhen-Hua Ling, Kai-Wei Chang, Nanyun Peng
Model editing methods, while effective at updating factual knowledge in Large Language Models, can significantly degrade their broader cognitive abilities like reasoning and question answering. A proposed regularization method, RECT, mitigates this degradation, preserving general abilities while maintaining over 94% of the intended editing performance.
Key insights
- Model editing, even with minimal parameter changes, substantially degrades LLMs' general abilities across diverse downstream tasks.
- The root cause of this degradation is that current editing methods induce excessive weight changes, leading to overfitting to the newly injected facts.
- The proposed RECT regularization method effectively preserves general abilities by preventing overly complex weight updates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.