Large language models (LLMs) possess a remarkable ability called in-context learning (ICL), allowing them to adapt to new tasks by processing examples within their prompts. Think of it like showing an LLM a few solved math problems before giving it a new one—it uses those examples as a guide. Retrieval-augmented generation (RAG) supercharges this by fetching relevant information from a database to include in the prompt. However, current methods for finding this information prioritize semantic similarity—how close the words are in meaning—rather than how useful the information is for actually solving the task. Imagine searching for help with algebra and getting geometry results—semantically similar, but not helpful! Researchers are addressing this with a new benchmark called ICLERB (In-Context Learning Embedding and Reranker Benchmark) that tests how well retrieval methods actually improve LLM accuracy on ICL tasks. They're also pioneering a reinforcement learning technique called RLRAIF (Reinforcement Learning-to-Rank from AI Feedback) to train retrieval models using direct feedback from the LLM. This tells the retriever exactly which information boosts performance, much like a student telling a tutor which explanations were most helpful. Early results are promising: smaller models trained with RLRAIF outperform much larger, state-of-the-art models on ICLERB, showing that picking the right data is more important than sheer size. This work signifies a critical shift in how we evaluate and enhance in-context learning, paving the way for LLMs that truly learn from experience and provide more accurate, contextually relevant results. While the current ICLERB focuses on multiple-choice questions, future research will expand to broader tasks and document types, further unlocking the potential of ICL and RAG for building more adaptable and knowledgeable LLMs.