DecisionRAG: Retriever-Free Retrieval with Decision Models That Navigate Documents

Forum topic · 小凯 · 2026-10-27

Summary

DecisionRAG is a new retrieval approach that replaces traditional retrievers with decision models built on pretrained language models. Instead of relying on lexical matching or similarity in fixed embedding spaces, it navigates a corpus's inherent structure—from broad topics down to individual passages—jointly evaluating candidate branches in a single forward pass without generating any tokens. Judging candidates in a shared context makes their relevance estimates comparable. Without task-specific training or a first-stage retriever, DecisionRAG outperforms dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial and 17.4 Hit@1 points on Natural Questions. Across four benchmarks, including a corpus of 2.68 million Wikipedia paragraphs, retrieval takes only 0.7–1.8 seconds per query. Paper by Chao Wan, Paul Jünger, Beixi Du, Bodong Liu, Navaneetha Krishnan et al., arXiv:2610.35868.

Paper Overview

Field: Machine Learning Authors: Chao Wan, Paul Jünger, Beixi Du, Bodong Liu, Navaneetha Krishnan et al. Published: 2026-10-26 arXiv: 2610.35868

Abstract

While 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.

The authors propose DecisionRAG, which brings the capabilities of pretrained language models to retrieval using decision models that jointly evaluate candidate branches in a single forward pass, without generating any tokens. Judging candidates in a shared context makes their relevance estimates comparable.

Like a person browsing a library, DecisionRAG navigates the corpus's inherent structure from broad topics down to individual passages, deciding at each level which branches are worth exploring.

Key Results

No task-specific training and no first-stage retriever required
Outperforms dense retrieval with an 8B reranker by 13.8 R@1 points on MultiDoc2Dial
Outperforms dense retrieval with an 8B reranker by 17.4 Hit@1 points on Natural Questions
Evaluated across four benchmarks, including a corpus of 2.68 million Wikipedia paragraphs
Retrieval takes just 0.7–1.8 seconds per query

---

*Auto-collected on 2026-10-28.*