DecisionRAG Retrieval Method Outperforms Dense Retrieval on MultiDoc2Dial and Natural Questions

Researchers proposed DecisionRAG, a retriever-free retrieval method that uses decision models to navigate corpus hierarchies from topics to passages in a single forward pass without generating tokens. Without task-specific training or a first-stage retriever, DecisionRAG outperformed dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions, retrieving from 2.68 million Wikipedia paragraphs in 0.7 to 1.8 seconds per query.
Event version 1.
Key insights
Retriever-free navigation
DecisionRAG replaces the first-stage retriever with a decision model that navigates a corpus hierarchy from broad topics to documents to passages. At each level, it scores all candidates in a shared context and keeps only the most promising branches.
Shared-context scoring gains
Scoring candidates together in one shared context improves R@1 on MultiDoc2Dial by 7.7 points and Hit@1 on Natural Questions by 6.5 points over judging each candidate separately.
Benchmark results
Without task-specific training or a first-stage retriever, DecisionRAG beats Qwen3-Embedding-8B dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions over 2.68 million Wikipedia paragraphs.
Cost and latency
Retrieval takes 0.7 to 1.8 seconds and 0.2 to 2.3 cents per query across four benchmarks. An LLM that generates navigation decisions token by token reaches similar accuracy at 34 to 145 times the latency.
Trained judge performance
A Qwen3-8B decision model trained on MultiDoc2Dial data reaches 60.7 R@1, 5.7 points above the strongest published task-trained retriever.