Tag
IAR is a three-stage post-training framework that internalizes structured document knowledge into language models for retrieval-free question answering, enhancing both domain-specific accuracy and general model performance.
This paper studies internalizing documents into LoRA adapters for closed-book QA and finds that once adapter capacity is sufficient, training data quality (especially answer conciseness) is the dominant factor, outperforming architectural changes and achieving higher accuracy than BM25-RAG.