| Retrieval-Augmented Generation: A Comprehensive Survey of Advanced Techniques and Their Evolution |
| کد مقاله : 1094-CYSP2025 (R1) |
| نویسندگان |
|
زهرا ابطحی فروشانی * دانشجوی کارشناسی ارشد مهندسی کامپیوتر، گرایش هوش مصنوعی و رباتیک، دانشکده مهندسی دانشکدگان فارابی دانشگاه تهران |
| چکیده مقاله |
| Retrieval-Augmented Generation (RAG) systems have significantly advanced the capabilities of Large Language Models (LLMs) by integrating external knowledge, thereby mitigating issues such as hallucinations and enhancing domain-specific applicability. While traditional RAG approaches, primarily relying on flat text representations, encounter limitations in complex query understanding, knowledge integration from distributed sources, and overall system efficiency, particularly for queries requiring multi-hop reasoning or global contextual understanding. This comprehensive review article synthesizes recent advancements in RAG, exploring novel architectural designs, improved retrieval mechanisms, and sophisticated generation strategies. We discuss frameworks that move beyond simple similarity-based retrieval to incorporate utility-oriented and compactness-oriented thoughts, and those that adapt LLM parameters through retrieval for zero-shot learning. Furthermore, we examine corrective RAG mechanisms, neurobiologically inspired long-term memory systems, and significant progress in multimodal RAG, including new benchmarks and structured visual understanding. The article highlights the evolution of RAG, its current capabilities across various modalities, and identifies ongoing challenges and future research directions, emphasizing the continuous pursuit of more robust efficient, and contextually faithful LLMs. |
| کلیدواژه ها |
| Large Language Models, Retrieval-Augmented Generation, Literature Survey. |
| وضعیت: پذیرفته شده برای ارائه شفاهی |