Researchers have proposed a new framework called SCOPE for compressing input prompts to large language models (LLMs) without requiring any training. SCOPE works by splitting the prompt into semantically coherent chunks and rewriting them to be more concise. This approach is designed to preserve critical information and text coherence while allowing for finer-grained control over compression ratios. The authors evaluated SCOPE on typical LLM applications, including question-answering and summarization, and found it outperformed existing selective compression methods in most settings, particularly at high compression ratios. --- Why it matters: This matters to engineers working with large language models because it provides a new approach to reducing the input context length without sacrificing generation quality. This could lead to significant reductions in inference cost and latency, making LLM applications more efficient and scalable. Source: https://arxiv.org/abs/2508.15813