Researchers have proposed a new framework for in-context search called LENS. It allows AI models to find relevant information in dynamic raw documents without pre-materializing the evidence space. Instead of relying on fixed chunking or persistent indexes, LENS uses an iterative process to select and update candidate units based on their relevance to the query. This approach enables more accurate and efficient retrieval of evidence, particularly for large-scale applications. The authors demonstrate the effectiveness of LENS through experiments on a controlled evaluation set and a fullwiki subset. --- Why it matters: This matters because it addresses a common challenge in AI research: handling dynamic and large-scale data sources without sacrificing performance or efficiency. Engineers can apply this framework to develop more robust and scalable search systems for various applications, such as question-answering models or information retrieval systems. Source: https://arxiv.org/abs/2608.16185