ROME
A knowledge editing method for large language models that locates weights associated with a target fact and applies a rank-one matrix update to modify them. The method is computationally efficient but may degrade general model capabilities when applied at scale.
Timeline (last 90 days)
Knowledge editing methods are mainly divided into three categories: direct parameter modification (e.g., ROME, MEMIT), external memory or retrieval augmentation (RAG), and fine-tuning.
Direct parameter modification methods achieve editing by locating weights associated with the target fact and performing matrix updates; they are fast but prone to degrading the model's general capabilities at scale.
Existing knowledge editing methods such as ROME and MEMIT typically assume that factual knowledge is concentrated in the MLP weights of certain layers and selectively modify those weights.
ROME和MEMIT是精准修改模型知识的知识编辑方法
All Facts (4)
Knowledge editing methods are mainly divided into three categories: direct parameter modification (e.g., ROME, MEMIT), external memory or retrieval augmentation (RAG), and fine-tuning.
50%UnverifiedDirect parameter modification methods achieve editing by locating weights associated with the target fact and performing matrix updates; they are fast but prone to degrading the model's general capabilities at scale.
50%UnverifiedExisting knowledge editing methods such as ROME and MEMIT typically assume that factual knowledge is concentrated in the MLP weights of certain layers and selectively modify those weights.
50%UnverifiedROME和MEMIT是精准修改模型知识的知识编辑方法
50%