Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

RA-ISF: Learning to Answer and Understand from Retrieval Augmentation via Iterative Self-Feedback

Yanming Liu, Xinyue Peng, Xuhong Zhang, Weihao Liu, Jianwei Yin, Jiannan Cao, Tianyu Du

TL;DR

RA-ISF presents a three-module iterative framework for retrieval-augmented question answering that jointly reasons over internal knowledge, retrieved passages, and decomposed sub-questions. By training lightweight sub-models for self-knowledge, relevance, and decomposition, and employing iterative refinement with a thresholded depth, RA-ISF reduces hallucinations and improves factual accuracy across diverse OpenQA benchmarks. Empirical results show large gains over retrieval baselines on GPT-3.5 and competitive gains with Llama-2, along with thorough ablations and efficiency analyses. The work highlights the practical impact of combining modular reasoning with retrieval in a controlled, iterative loop for robust knowledge-intensive QA.

Abstract

Large language models (LLMs) demonstrate exceptional performance in numerous tasks but still heavily rely on knowledge stored in their parameters. Moreover, updating this knowledge incurs high training costs. Retrieval-augmented generation (RAG) methods address this issue by integrating external knowledge. The model can answer questions it couldn't previously by retrieving knowledge relevant to the query. This approach improves performance in certain scenarios for specific tasks. However, if irrelevant texts are retrieved, it may impair model performance. In this paper, we propose Retrieval Augmented Iterative Self-Feedback (RA-ISF), a framework that iteratively decomposes tasks and processes them in three submodules to enhance the model's problem-solving capabilities. Experiments show that our method outperforms existing benchmarks, performing well on models like GPT3.5, Llama2, significantly enhancing factual reasoning capabilities and reducing hallucinations.

RA-ISF: Learning to Answer and Understand from Retrieval Augmentation via Iterative Self-Feedback

TL;DR

RA-ISF presents a three-module iterative framework for retrieval-augmented question answering that jointly reasons over internal knowledge, retrieved passages, and decomposed sub-questions. By training lightweight sub-models for self-knowledge, relevance, and decomposition, and employing iterative refinement with a thresholded depth, RA-ISF reduces hallucinations and improves factual accuracy across diverse OpenQA benchmarks. Empirical results show large gains over retrieval baselines on GPT-3.5 and competitive gains with Llama-2, along with thorough ablations and efficiency analyses. The work highlights the practical impact of combining modular reasoning with retrieval in a controlled, iterative loop for robust knowledge-intensive QA.

Abstract

Large language models (LLMs) demonstrate exceptional performance in numerous tasks but still heavily rely on knowledge stored in their parameters. Moreover, updating this knowledge incurs high training costs. Retrieval-augmented generation (RAG) methods address this issue by integrating external knowledge. The model can answer questions it couldn't previously by retrieving knowledge relevant to the query. This approach improves performance in certain scenarios for specific tasks. However, if irrelevant texts are retrieved, it may impair model performance. In this paper, we propose Retrieval Augmented Iterative Self-Feedback (RA-ISF), a framework that iteratively decomposes tasks and processes them in three submodules to enhance the model's problem-solving capabilities. Experiments show that our method outperforms existing benchmarks, performing well on models like GPT3.5, Llama2, significantly enhancing factual reasoning capabilities and reducing hallucinations.
Paper Structure (32 sections, 5 equations, 4 figures, 7 tables, 1 algorithm)

This paper contains 32 sections, 5 equations, 4 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Retrieval Augmented Language Model
  4. Task Decomposition
  5. Methodology
  6. Overview
  7. RA-ISF Training
  8. RA-ISF Inference
  9. Experimental Setup
  10. Datasets
  11. Models
  12. Retriever and Corpus
  13. Baselines
  14. Implementation Details
  15. Experiment Results
  16. ...and 17 more sections

Figures (4)

  • Figure 1: Overview of RA-ISF. It consists of three sub-modules: self-knowledge, passage relevance, and question decomposition.
  • Figure 2: Question accuracy on the NQ dataset with the growth of the iteration in question decomposition $D_{th}$.
  • Figure 3: Trend of question accuracy on the NQ dataset with the growth of the iteration in question decomposition $k$.
  • Figure 4: Trend of question accuracy on the NQ and TriviaQA dataset with the growth of the iteration in question decomposition $k$.