Frameworks for Robustness Evaluation and Enhancement of Retrieval-Augmented Generation

• 1、School of Cyberspace Science And Technology, Beijing Jiaotong University ,Beijing 100044, China

Abstract：Retrieval-Augmented Generation (RAG) systems enhance the factual accuracy of Large Language Model (LLM) by integrating external knowledge bases. However, insufficient robustness in these systems can lead to the generation of non-factual hallucinated information by the LLM. To address the robustness deficiencies in RAG, this research explores the problem from the dual dimensions of evaluation and enhancement, proposing a comprehensive framework for improving RAG system reliability. At the evaluation level, we introduce the RAISP (Robustness and Availability Assessment via Internal Sensitive Knowledge Profiling) framework. This framework simulates retrieval biases in controlled settings via a gradient-based suffix optimization strategy, enabling a quantitative analysis of a RAG system\'s vulnerability to minor perturbations. At the enhancement level, we propose the PGDRAG (Purification and Geometric Denoising RAG) framework, which constructs a three-stage pipeline comprising "Query Purification, Geometric Denoising, and Hierarchical Verification." Experimental results on multiple datasets demonstrate that the RAISP evaluation framework can effectively expose systemic vulnerabilities, while the PGDRAG defense framework significantly reduces the hallucination rate under adversarial interference. This work provides technical support for the reliable deployment of RAG systems in high-stakes domains.

Keywords： Large Language Model Retrieval-Augmented Generation Robustness Enhancement Hallucination