The field of large language models (LLMs) is moving towards incorporating external knowledge to improve performance, particularly in question-answering and reasoning tasks. Retrieval-augmented generation (RAG) has emerged as a key approach, focusing on balancing internal knowledge with external information retrieval. Recent developments highlight the need for dynamic and adaptive methods to manage knowledge conflicts, ambiguity, and noise in retrieved information. Noteworthy papers include RAG-VR, which improves answer accuracy by 17.9%-41.8% and reduces latency by 34.5%-47.3% in 3D question-answering, and ARise, which integrates risk assessment with dynamic RAG to achieve significant improvements in knowledge-augmented reasoning. Additionally, ACoRN introduces a novel training approach to enhance compressor robustness against retrieval noise, and MADAM-RAG proposes a multi-agent debate framework to handle conflicting evidence and misinformation. These innovative approaches demonstrate the potential for RAG to advance the field of LLMs and improve their reliability in real-world applications.
Advancements in Retrieval-Augmented Generation for Large Language Models
Sources
RAG-VR: Leveraging Retrieval-Augmented Generation for 3D Question Answering in VR Environments
CCSK:Cognitive Convection of Self-Knowledge Based Retrieval Augmentation for Large Language Models
HeteRAG: A Heterogeneous Retrieval-augmented Generation Framework with Decoupled Knowledge Representations
ARise: Towards Knowledge-Augmented Reasoning via Risk-Adaptive Search
Leveraging Agency in Virtual Reality to Enable Situated Learning
ACoRN: Noise-Robust Abstractive Compression in Retrieval-Augmented Language Models
Accommodate Knowledge Conflicts in Retrieval-augmented LLMs: Towards Reliable Response Generation in the Wild
Retrieval-Augmented Generation with Conflicting Evidence