Retrieved In-Context Principles from Previous Mistakes
Researchers have developed a new framework called Retrieved In-Context Principles (RICP) to improve the performance of large language models. RICP utilizes a teacher-student model where the teacher analyzes the student's mistakes to generate insights and principles for error prevention. These principles are then customized for specific questions by retrieving the most relevant mistakes, enhancing guidance without requiring teacher intervention during inference. Experiments on seven reasoning benchmarks show that RICP effectively boosts performance across various prompting strategies. AI
IMPACT Enhances LLM reasoning capabilities by providing a novel method for learning from and preventing errors.