Researchers built DecisionRAG, a retrieval method that lets a language model navigate documents like a person browsing a library. It judges which topic branches are worth exploring, no separate retriever or token generation needed. In early benchmarks, it beat standard dense retrieval systems on multiple question-answering tests.
Today, if a search system's first step misses the right document, even the smartest AI can't recover the answer. DecisionRAG attacks that weak link directly. It outperformed a strong baseline by 13.8 points on dialogue search and 17.4 on natural questions, while scanning 2.68 million Wikipedia paragraphs in under two seconds per query.
✦ Better finding, not just better answering, could make AI helpers genuinely more reliable.