Researchers developed Inject, Align, and Recover (IAR), a three-stage post-training framework that lets language models answer questions about specific document collections without needing retrieval at inference time. The method converts source documents into structured knowledge through specialized training objectives in an injection phase, adapts the model with QA supervision, then merges it back with the base model to preserve general capabilities. Across multiple model families and datasets, IAR outperformed conventional supervised fine-tuning on both domain-specific question answering and general benchmarks.
