Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization
IAR turns a fixed document corpus into parametric knowledge through separate injection, QA alignment, and recovery stages. It improved domain QA while retaining more general capability than vanilla SFT.
IAR separates retrieval-free document internalization into **Inject, Align, and Recover**. Documents become continuation, rewrite, and reconstruction objectives; answer-only QA follows; then the adapted model is merged with the base instruction model.
For bounded corpora that must work without runtime retrieval, separate knowledge ingestion from answer behavior and general-capability recovery. The method improved all four reported metrics over Vanilla SFT in **7 of 8 settings**.
IAR separates retrieval-free document internalization into **Inject, Align, and Recover**. Documents become continuation, rewrite, and reconstruction objectives; answer-only QA follows; then the adapted model is merged with the base instruction model. For bounded corpora that must work without runtime retrieval, separate knowledge ingestion from answer behavior and general-capability recovery. The method improved all four reported metrics over Vanilla SFT in **7 of 8 settings**. Average gains were **3.6 percentage points** in domain QA and **12.1 points** in mean general performance. LoRA and FAPM still won some individual general metrics, so IAR is a frontier tradeoff rather than a universal winner.
This supplies a concrete staged method for the earlier claim that private-corpus training needs study objectives rather than a single fine-tuning pass, while adding an explicit recovery step for overwritten general capability. It also narrows the context-architecture choice: bounded, stable corpora may be internalized without runtime retrieval, whereas changing or provenance-sensitive knowledge still favors external context systems.