LENS drops persistent indexes for query-time evidence search
TL;DR
- On a 500-question controlled benchmark, LENS reached 62.4% exact match with 84.8% evidence recall versus 65.2% EM and 50.4% recall for a ReAct baseline.
- The framework is index-free: rather than chunking and embedding a corpus, it iteratively probes raw documents with an LLM relevance oracle under a controlled budget.
- On a 150-question fullwiki Wikipedia subset, LENS and ReAct nearly tied on EM (43.3% vs 42.7%) but LENS grounded 84.0% of answers in retrieved evidence versus 70.7%.
On a controlled 500-question benchmark, an index-free framework called LENS answered 62.4% of questions correctly against 65.2% for a ReAct-style baseline. It grounded those answers in real retrieved evidence at 84.8% recall, versus 50.4% for ReAct.
The preprint from Xingjun Wang and colleagues, posted August 17, formulates search as "Budgeted Evidence Localization over a latent evidence space induced by dynamic raw documents." Rather than chunking and embedding a corpus in advance, LENS "maintains a query-conditioned belief over candidate units, iteratively selecting candidates via complementary lexical, local, and exploratory proposal policies," with an LLM acting as the relevance oracle inside a budget.
On a 150-question fullwiki subset over the raw Wikipedia dump, LENS and ReAct were nearly tied on answer quality (43.3% vs 42.7% EM), but LENS grounded 84.0% of its answers in retrieved text against 70.7%. The authors write that LENS is "query-ready after corpus changes" and needs "no preprocessing or persistent index."
Two researchers we track posted the paper the same day.
Shared on Bluesky by 2 AI experts
Originally reported by arxiv.org
Read the original article →Original headline: LENS: In-Context Search via Latent Evidence Exploration over Dynamic Raw Documents