DecisionRAG: Decision Models Beat 8B-Reranker Dense Retrieval by Up to 17.4 Points
kalyan_kpl · x · 2026-10-30
The paper introduces DecisionRAG, which replaces dense embedding models with Decision Models for retrieval. Like a person browsing a library, it navigates the corpus's inherent structure from broad topics down to individual passages, deciding at each level which branches to explore. Without task-specific training or a first-stage retriever, it outperforms dense retrieval with an 8B reranker across four benchmarks: +13.8 R@1 on MultiDoc2Dial and +17.4 Hit@1 on Natural Questions.
More from Research
- Gaussian head on flow policies: 0.02% of data, similar success, up to 2x faster inference — francoisfleuret · 2026-10-30
- Northwestern's Hullman: Verification is not knowledge — AI science needs humans who can answer for results — JessicaHullman · 2026-10-30
- Humansand launches &Tutor, a tutor model trained by teaching a human-like learner model — techNmak · 2026-10-30
- CATT benchmark shows attribution methods fail to trace model answers to training data — cindy_x_wu · 2026-10-30
- Launch: personalized cancer vaccine platform to manufacture 100+ mRNA vaccines — jajoosam · 2026-10-30
- Benchmark sweep shows greedy decoding tanks coding accuracy: 70% at temp 0 vs 82.7% at temp 1.0 — Brief-Tap-6616 · 2026-10-30