The paper introduces Learning Principles (LEAP), a novel approach to in-context learning (ICL) that enhances the performance of large language models (LLMs) on downstream tasks. Unlike traditional ICL methods that only learn from correct input-output pairs, LEAP induces the model to make mistakes on a few given examples, reflects on these mistakes, and learns explicit task-specific principles. These principles are then used to improve the model's performance on unseen test questions. The approach is evaluated on various benchmarks, including multi-hop question answering, textual reasoning, and mathematical reasoning, showing significant improvements over standard ICL methods using strong LLMs such as GPT-3.5-turbo, GPT-4, GPT-4-turbo, and Claude-2.1. LEAP requires no additional input or examples beyond the standard few-shot prompting settings and demonstrates the effectiveness of learning from mistakes in improving model performance.The paper introduces Learning Principles (LEAP), a novel approach to in-context learning (ICL) that enhances the performance of large language models (LLMs) on downstream tasks. Unlike traditional ICL methods that only learn from correct input-output pairs, LEAP induces the model to make mistakes on a few given examples, reflects on these mistakes, and learns explicit task-specific principles. These principles are then used to improve the model's performance on unseen test questions. The approach is evaluated on various benchmarks, including multi-hop question answering, textual reasoning, and mathematical reasoning, showing significant improvements over standard ICL methods using strong LLMs such as GPT-3.5-turbo, GPT-4, GPT-4-turbo, and Claude-2.1. LEAP requires no additional input or examples beyond the standard few-shot prompting settings and demonstrates the effectiveness of learning from mistakes in improving model performance.