Model Editing

Model editing studies how to update a trained model’s knowledge or behavior without retraining it from scratch, for tasks such as correcting facts, removing sensitive information, personalizing responses, and improving safety. Common approaches locate influential layers or neurons and apply low-rank or closed-form weight changes, while others use adapters, external memories, routing, or retrieval to store and apply edits. Researchers evaluate whether edits work on paraphrases and related questions, stay local to the target, preserve general capabilities, and remain stable across sequential updates, including in multimodal and generative models. A central challenge is balancing useful generalization with protection against unintended changes, forgetting, and degradation after further fine-tuning.