Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Decide Then Retrieve: A Training-Free Framework with Uncertainty-Guided Triggering and Dual-Path Retrieval

Wang Chen, Guanqiang Qi, Weikang Li, Yang Li, Deguo Xia, Jizhou Huang

TL;DR

Decide Then Retrieve (DTR) tackles the brittleness of retrieval-augmented generation by making retrieval activation and evidence selection adaptive and training-free. It combines uncertainty-guided triggering with a dual-path retrieval strategy and adaptive information selection to reduce noise and improve answer quality, especially on sparse or ambiguous queries. Across five open-domain QA benchmarks and multiple model scales, DTR consistently outperforms standard RAG and retrieval-enhanced baselines, while reducing unnecessary retrievals. The approach is model-agnostic, robust to different retrievers, and demonstrates practical impact for deploying adaptive RAG systems in real-world QA tasks.

Abstract

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external knowledge, but existing approaches indiscriminately trigger retrieval and rely on single-path evidence construction, often introducing noise and limiting performance gains. In this work, we propose Decide Then Retrieve (DTR), a training-free framework that adaptively determines when retrieval is necessary and how external information should be selected. DTR leverages generation uncertainty to guide retrieval triggering and introduces a dual-path retrieval mechanism with adaptive information selection to better handle sparse and ambiguous queries. Extensive experiments across five open-domain QA benchmarks, multiple model scales, and different retrievers demonstrate that DTR consistently improves EM and F1 over standard RAG and strong retrieval-enhanced baselines, while reducing unnecessary retrievals. The code and data used in this paper are available at https://github.com/ChenWangHKU/DTR.

Decide Then Retrieve: A Training-Free Framework with Uncertainty-Guided Triggering and Dual-Path Retrieval

TL;DR

Decide Then Retrieve (DTR) tackles the brittleness of retrieval-augmented generation by making retrieval activation and evidence selection adaptive and training-free. It combines uncertainty-guided triggering with a dual-path retrieval strategy and adaptive information selection to reduce noise and improve answer quality, especially on sparse or ambiguous queries. Across five open-domain QA benchmarks and multiple model scales, DTR consistently outperforms standard RAG and retrieval-enhanced baselines, while reducing unnecessary retrievals. The approach is model-agnostic, robust to different retrievers, and demonstrates practical impact for deploying adaptive RAG systems in real-world QA tasks.

Abstract

Retrieval-augmented generation (RAG) enhances large language models (LLMs) by incorporating external knowledge, but existing approaches indiscriminately trigger retrieval and rely on single-path evidence construction, often introducing noise and limiting performance gains. In this work, we propose Decide Then Retrieve (DTR), a training-free framework that adaptively determines when retrieval is necessary and how external information should be selected. DTR leverages generation uncertainty to guide retrieval triggering and introduces a dual-path retrieval mechanism with adaptive information selection to better handle sparse and ambiguous queries. Extensive experiments across five open-domain QA benchmarks, multiple model scales, and different retrievers demonstrate that DTR consistently improves EM and F1 over standard RAG and strong retrieval-enhanced baselines, while reducing unnecessary retrievals. The code and data used in this paper are available at https://github.com/ChenWangHKU/DTR.
Paper Structure (49 sections, 7 equations, 19 figures, 7 tables, 1 algorithm)

This paper contains 49 sections, 7 equations, 19 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Adaptive retrieval
  4. Query augmentation
  5. Preliminaries
  6. Problem setup
  7. Accuracy analysis of RAG
  8. Decide Then Retrieve
  9. Uncertainty--Guided Triggering
  10. Dual-path retrieval
  11. Adaptive information selection.
  12. Experiments
  13. Experiment setup
  14. Evaluation datasets.
  15. Baselines.
  16. ...and 34 more sections

Figures (19)

  • Figure 1: Overview of the proposed decide then retrieve (DTR) framework. (a) DTR can adaptively determine whether to retrieve and how to select external information. (b) DTR guides whether to activate retrievals based on the uncertainty score. (c) DTR adaptively selects effective information based on the dual-path retrieval for the final generation.
  • Figure 2: Overview of RAG accuracy analysis.
  • Figure 3: Generation Uncertainty vs. Parametric Accuracy and Query Coverage. Qwen2.5 series models are used as the generators.
  • Figure 4: Illustration of (a) dual-path retrieval mechanism and (b) similarity calculation of retrievals.
  • Figure 5: Uncertainty scaling results across various model sizes and different retrieval mechanisms.
  • ...and 14 more figures