Large Language Models (LLMs) possess remarkable abilities, but their full potential is often limited by practical constraints. One such constraint arises in in-context learning (ICL), where LLMs learn from examples provided directly within the input prompt. While effective, ICL can lead to excessively long prompts, straining hardware resources. Furthermore, not all demonstrations are created equal; some contribute more to the learning process than others. How can we make ICL more efficient and effective? Researchers have explored various techniques, such as prompt pruning and soft prompts, to compress demonstrations and reduce input length. They've also developed methods for selecting the most impactful demonstrations. However, these approaches often involve separate modules, adding complexity and overhead. A new framework called UniICL offers a more streamlined solution. UniICL unifies demonstration selection, compression, and response generation within a single, frozen LLM. It leverages the LLM's inherent understanding to compress different demonstrations independently into compact representations called virtual tokens. These virtual tokens not only replace the original, lengthy demonstrations but also serve as the basis for selecting relevant examples. By combining these processes, UniICL significantly reduces the input size, enabling LLMs to handle more demonstrations and potentially unlock higher performance. Experiments across various tasks, from linguistic acceptability to text summarization, show that UniICL effectively compresses demonstrations by up to 12x while maintaining or even improving performance. This allows for scaling from 4-shot to 64-shot ICL within limited hardware constraints. UniICL also introduces a "Demonstration Bank" to cache compressed tokens, further boosting efficiency by avoiding redundant computations. While promising, UniICL has limitations. It currently focuses on basic ICL, leaving more advanced prompting methods like Retrieval Augmented Generation (RAG) and Chain-of-Thought (CoT) for future exploration. Furthermore, the research primarily uses 7-billion parameter LLMs, and scaling to larger models could reveal further insights. Despite these limitations, UniICL represents a significant step towards more efficient and powerful in-context learning, paving the way for LLMs to tackle even more complex tasks.