DecisionRAG retrieves passages with token-free decision models
Original title: Retriever-Free Retrieval: Navigating Documents with Decision Models
Without task-specific training, it beats dense retrieval plus an 8B reranker by 13.8 R@1 on MultiDoc2Dial.
Researchers propose DecisionRAG, which uses pretrained language models as decision models to score candidate document branches in one forward pass without generating tokens. It walks corpus structure from broad topics down to passages and needs no first-stage retriever. Across four benchmarks, including 2.68 million Wikipedia paragraphs, retrieval takes 0.7-1.8 seconds per query. Reported gains are 13.8 R@1 on MultiDoc2Dial and 17.4 Hit@1 on Natural Questions.
- Shared-context scoring is intended to make relevance estimates comparable.
- No task-specific training or first-stage retriever is used.
- The largest cited corpus has 2.68 million Wikipedia paragraphs.
- Reported latency is 0.7-1.8 seconds per query.
From the original article
1 min read at arxiv.orgWhile language models have become dramatically more capable, retrieval still largely relies on lexical matching or similarity in fixed embedding spaces. This creates a capability asymmetry: what the retriever fails to find, even the strongest generator cannot recover.
By Chao Wan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.