LENS: Index-Free Framework for In-Context Search Over Dynamic Documents

A recent preprint on arXiv presents LENS (Latent Evidence Exploration and Search), an innovative framework designed for in-context search within ever-changing raw-document collections without the need for indexing. The paper, identified as arXiv:2608.16185v2, tackles the difficulties encountered by LLM agents when responding to queries from documents that may alter prior to preprocessing, with relevant evidence being dependent on the specific query. Traditional retrieval-augmented methods rely on fixed chunking, embeddings, or persistent indexes, which can be expensive and outdated. In contrast, LENS approaches in-context search as Budgeted Evidence Localization within a latent evidence space formed by dynamic documents. It maintains a belief conditioned on queries and iteratively selects candidates, refining this belief through an LLM relevance oracle to focus on high-posterior areas, thus providing a more flexible solution for dynamic collections.