Optimizing White Noise for Stability and Strategic Performance Enhancement of LLM-RAG Systems

Retrieval-Augmented Generation (RAG) has emerged as a technique for improving the performance of language models (LLMs), where the inclusion of noisy documents is unavoidable. While prior studies have largely treated noise as detrimental and focused on its removal, this study demonstrates that certain types and combinations of noisy documents can improve LLM-based question answering performance. We categorize noisy documents into six types and systematically evaluate their individual and combined effects within an LLM-RAG framework. To analyze the mechanisms underlying performance changes, we employ quantitative metrics that capture improvement and degradation. We define beneficial noise as White Noise and harmful noise as Black Noise, framing noise along a brightness continuum. Furthermore, we propose a Pure White Noise selection method that provides practical guidance for leveraging beneficial noise to enhance the robustness of RAG systems. Our results suggest that noise can be strategically utilized as a resource for robust RAG design.