Researchers have proposed a new method called SKILL-RAG to improve the performance of large language models on knowledge-intensive tasks. The approach involves leveraging the model's internal knowledge, or 'self-knowledge', to filter out irrelevant content from external sources and focus on useful information. This is achieved through a reinforcement learning-based training framework that elicits self-knowledge at sentence-level granularity. Experiments show that SKILL-RAG improves generation quality and reduces the number of input documents required for answering questions. --- Why it matters: This matters to AI researchers because it addresses a key challenge in Retrieval-Augmented Generation: incorporating external knowledge without introducing hallucinations. By leveraging self-knowledge, SKILL-RAG can improve the performance of large language models on knowledge-intensive tasks. Source: https://arxiv.org/abs/2509.20377