Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM-Independent Adaptive RAG: Let the Question Speak for Itself

Maria Marina, Nikolay Ivanov, Sergey Pletenev, Mikhail Salnikov, Daria Galimzianova, Nikita Krayko, Vasily Konovalov, Alexander Panchenko, Viktor Moskvoretskii

TL;DR

The paper tackles the high cost and potential misinformation risk of retrieval-augmented QA by proposing an LLM-independent adaptive retrieval approach driven by external information. It introduces seven feature groups totaling 27 features that precompute retrieval decisions without relying on LLM uncertainty, and evaluates them across six QA datasets. Results show that external features can match or exceed uncertainty-based adaptive methods in QA quality while substantially reducing LM calls, highlighting significant efficiency gains. This work demonstrates that external information can substitute or complement internal uncertainty signals, enabling scalable and efficient RAG deployments, especially for complex questions.

Abstract

Large Language Models~(LLMs) are prone to hallucinations, and Retrieval-Augmented Generation (RAG) helps mitigate this, but at a high computational cost while risking misinformation. Adaptive retrieval aims to retrieve only when necessary, but existing approaches rely on LLM-based uncertainty estimation, which remain inefficient and impractical. In this study, we introduce lightweight LLM-independent adaptive retrieval methods based on external information. We investigated 27 features, organized into 7 groups, and their hybrid combinations. We evaluated these methods on 6 QA datasets, assessing the QA performance and efficiency. The results show that our approach matches the performance of complex LLM-based methods while achieving significant efficiency gains, demonstrating the potential of external information for adaptive retrieval.

LLM-Independent Adaptive RAG: Let the Question Speak for Itself

TL;DR

The paper tackles the high cost and potential misinformation risk of retrieval-augmented QA by proposing an LLM-independent adaptive retrieval approach driven by external information. It introduces seven feature groups totaling 27 features that precompute retrieval decisions without relying on LLM uncertainty, and evaluates them across six QA datasets. Results show that external features can match or exceed uncertainty-based adaptive methods in QA quality while substantially reducing LM calls, highlighting significant efficiency gains. This work demonstrates that external information can substitute or complement internal uncertainty signals, enabling scalable and efficient RAG deployments, especially for complex questions.

Abstract

Large Language Models~(LLMs) are prone to hallucinations, and Retrieval-Augmented Generation (RAG) helps mitigate this, but at a high computational cost while risking misinformation. Adaptive retrieval aims to retrieve only when necessary, but existing approaches rely on LLM-based uncertainty estimation, which remain inefficient and impractical. In this study, we introduce lightweight LLM-independent adaptive retrieval methods based on external information. We investigated 27 features, organized into 7 groups, and their hybrid combinations. We evaluated these methods on 6 QA datasets, assessing the QA performance and efficiency. The results show that our approach matches the performance of complex LLM-based methods while achieving significant efficiency gains, demonstrating the potential of external information for adaptive retrieval.
Paper Structure (25 sections, 2 equations, 5 figures, 2 tables)

This paper contains 25 sections, 2 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methods
  4. External Information Methods
  5. Experimental Setup
  6. Implementation Details
  7. Datasets
  8. Evaluation
  9. Results
  10. Downstream Performance
  11. Efficiency Performance
  12. Features Reciprocity
  13. Conclusion
  14. External Methods
  15. Graph
  16. ...and 10 more sections

Figures (5)

  • Figure 1: PFLOPs-Inaccuracy trade-off for proposed features vs the most efficient alternative adaptive retrieval methods for the NQ dataset. Radius of the points is proportional to the number of LLM calls. Green dotted line indicate Always RAG approach.
  • Figure 2: Feature importances for one of the best algorithms for only external features vs all features for TriviaQA (simple) and Musique (complex) datasets.
  • Figure 3: Heatmap of different groups of features for TriviaQA and 2WikiMultiHopQA (2wiki) datasets. Upper right triangle states for the absolute correlations on the TriviaQA, while down left states for the absolute correlations on the 2WikiMultiHopQA
  • Figure 4: Feature importances for one of the best algorithms for only external features vs all features for NQ, TriviaQA (simple) and HotpotQA, Musique (complex) datasets.
  • Figure 5: Absolute correlation of features from different groups of external features with class label