Researchers have developed LENS (Latent Evidence Exploration and Search), a novel framework for in-context search over dynamic raw documents. Unlike traditional retrieval-augmented methods that rely on pre-materialized evidence, LENS operates without a persistent index. It maintains a query-conditioned belief over candidate evidence units, iteratively refining this belief using an LLM relevance oracle and complementary proposal policies within a controlled budget. This approach allows LENS to adapt to document changes and provides stronger supporting-fact localization and answer grounding compared to existing baselines. AI
IMPACT This index-free approach could streamline LLM agent operations by reducing reliance on costly and stale evidence indexes.
RANK_REASON The cluster describes a new research paper detailing a novel framework for in-context search.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
