# Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization

Source: [arXiv](https://arxiv.org/abs/2608.20281v1)  
Feed7 permalink: https://feed7.dev/p/2608-20281v1-0radlbg  
Published: 2026-08-20T17:14:24.000Z  
Trust: Needs Review (needs_review)

## Why Included

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.

## Source Summary

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.

## Practical Implication

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**.

## Agent-Ready Context

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.

## Connected Context

Feed7 judgment across 525 accumulated Signals:

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.

- [Scaling Compute on Context — Jack Morris, Engram](https://feed7.dev/p/scaling-compute-on-context-jack-morris-engram-1pxr9bt) — IAR operationalizes this candidate’s proposed synthetic study tasks and directly addresses its warning about overwritten prior knowledge through base-model merging.
- [The Rise of CaaS: Context-as-a-Service for Agentic AI — Omer Primor, Bright Data](https://feed7.dev/p/the-rise-of-caas-context-as-a-service-for-agentic-ai-omer-primor-bright-0bfgium) — They define complementary choices for stable repeated knowledge: owning retrieval infrastructure versus internalizing a bounded corpus into model parameters.
- [virgiliojr94/book-to-skill](https://feed7.dev/p/book-to-skill-1av16sr) — Both reduce repeated document loading, but book-to-skill preserves segmented references as an external artifact while IAR moves knowledge into model parameters and therefore changes provenance and update tradeoffs.

## Context Map

- Layer: model
- Domains: research, data
- Topics: retrieval, context-engineering

## Uncertainty

- 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.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.