Grounded Decoding: Retrieval-Anchored Probability Fusion for Faithful RAG
Researchers have developed a new training-free decoding framework called Grounded Decoding to enhance factual consistency in retrieval-augmented generation (RAG) systems. This method fuses probability distributions from both full RAG and retrieval-only conditions to ensure that language models prioritize external evidence. Experiments show improvements in factual accuracy and citation quality without sacrificing fluency, offering an efficient alternative to existing RAG decoding techniques. AI
IMPACT Enhances factual accuracy in RAG systems, potentially improving reliability for AI applications relying on external data.