Key Points
1. Retrieval-Augmented Generation (RAG) can be undermined by imperfect retrieval, which may introduce irrelevant, misleading, or even malicious information.
2. Imperfect retrieval augmentation is widespread even with adept real-world search engines - roughly 70% of retrieved passages do not directly contain true answers.
3. Knowledge conflicts between the LLM's internal knowledge and external retrieved knowledge is a key bottleneck that needs to be addressed.
4. As t u t e R AG is proposed as a novel RAG approach that aims to address conflicts between internal and external knowledge.
5. As t u t e R AG adaptively elicits information from the LLM's internal knowledge, iteratively consolidates internal and external knowledge with source-awareness, and finalizes the answer based on information reliability.
6. Experiments on Gemini and Claude demonstrate the superior performance of As t u t e R AG compared to previous robustness-enhanced RAG methods.
7. Notably, As t u t e R AG is the only RAG approach that matches or exceeds the performance of LLMs without RAG under worst-case scenarios.
8. Further analysis reveals As t u t e R AG effectively resolves knowledge conflicts, improving the reliability and trustworthiness of RAG systems.
9. The core contributions are: 1) analysis of RAG under realistic conditions, 2) the As t u t e R AG approach to address internal-external knowledge conflicts, and 3) experiments demonstrating the effectiveness of As t u t e R AG.
Summary
The paper explores the challenges and limitations of Retrieval-Augmented Generation (RAG) models, which aim to address the inherent knowledge limitations of large language models (LLMs) by integrating external knowledge from sources like corpora or knowledge bases.
The key findings are: 1. Imperfect retrieval is a widespread issue in real-world RAG systems, with around 70% of retrieved passages not directly containing the true answer. This leads to the RAG model's performance falling behind LLMs without retrieval augmentation, especially under low retrieval quality. 2. The paper identifies knowledge conflicts between the LLM's internal knowledge and the external knowledge retrieved as a key bottleneck that undermines the effectiveness of RAG. The LLM often struggles to reliably consolidate conflicting information from these two sources. To address these challenges, the paper proposes a novel approach called AstuteRAG. The key components of AstuteRAG are: 1. Adaptive generation of internal knowledge: AstuteRAG first elicits relevant information from the LLM's internal knowledge, guided by principles of accuracy and relevance, to supplement the potentially noisy retrieved passages. 2. Iterative source-aware knowledge consolidation: AstuteRAG then employs the LLM to iteratively consolidate the internal and external knowledge sources, identifying consistent information, detecting conflicts, and filtering out irrelevant information in a source-aware manner. 3. Answer finalization: Finally, AstuteRAG compares the reliability of answers generated from the consolidated knowledge groups to determine the most reliable final answer.
Experimental Results
Experiments on diverse datasets and with different LLMs, including Gemini and Claude, demonstrate that AstuteRAG significantly outperforms previous robustness-enhanced RAG methods. Notably, AstuteRAG is the only RAG approach that matches or exceeds the performance of LLMs without retrieval augmentation even under the worst-case scenario where all retrieved passages are unhelpful. Further analysis reveals the effectiveness of AstuteRAG in resolving knowledge conflicts, improving the reliability and trustworthiness of RAG systems.
Reference: https://arxiv.org/abs/2410.07176