The paper introduces the *Learning to Edit* (LTE) framework, which aims to enhance the ability of large language models (LLMs) to efficiently and effectively update their knowledge without negatively impacting overall performance. Traditional knowledge editing methods often rely on memorizing updated knowledge, which hinders LLMs from integrating new information with their existing knowledge. In contrast, LTE focuses on teaching LLMs to dynamically apply updated knowledge to input questions, inspired by the principle of "teach a man to fish." The LTE framework consists of two phases: 1. **Alignment Phase**: This phase involves fine-tuning LLMs on a curated parallel dataset to ensure they can make reliable, in-scope edits while preserving out-of-scope information and maintaining linguistic proficiency. 2. **Inference Phase**: This phase employs a retrieval-based mechanism to retrieve relevant updated knowledge from a memory bank for real-time and mass knowledge editing. The paper evaluates LTE against seven advanced baselines across four popular knowledge editing benchmarks and two LLM architectures. The results demonstrate that LTE outperforms existing methods in terms of overall performance, robustness in batch and sequential editing, minimal interference on general tasks, and rapid editing speeds. LTE establishes a new state-of-the-art in knowledge editing tasks, showcasing its effectiveness and efficiency in integrating new knowledge into LLMs' outputs.The paper introduces the *Learning to Edit* (LTE) framework, which aims to enhance the ability of large language models (LLMs) to efficiently and effectively update their knowledge without negatively impacting overall performance. Traditional knowledge editing methods often rely on memorizing updated knowledge, which hinders LLMs from integrating new information with their existing knowledge. In contrast, LTE focuses on teaching LLMs to dynamically apply updated knowledge to input questions, inspired by the principle of "teach a man to fish." The LTE framework consists of two phases: 1. **Alignment Phase**: This phase involves fine-tuning LLMs on a curated parallel dataset to ensure they can make reliable, in-scope edits while preserving out-of-scope information and maintaining linguistic proficiency. 2. **Inference Phase**: This phase employs a retrieval-based mechanism to retrieve relevant updated knowledge from a memory bank for real-time and mass knowledge editing. The paper evaluates LTE against seven advanced baselines across four popular knowledge editing benchmarks and two LLM architectures. The results demonstrate that LTE outperforms existing methods in terms of overall performance, robustness in batch and sequential editing, minimal interference on general tasks, and rapid editing speeds. LTE establishes a new state-of-the-art in knowledge editing tasks, showcasing its effectiveness and efficiency in integrating new knowledge into LLMs' outputs.