AAI News Hub
ResearchThu, August 20, 2026·1d ago

Researchers propose IAR for retrieval-free document QA

The staged post-training method aims to internalize fixed document collections without inference-time retrieval.

Why it matters

The work targets a practical limitation of LLMs: answering from fixed document sets without relying on retrieval at inference time. If reproducible, staged internalization could offer an alternative path for document-specific assistants where retrieval is unavailable or undesirable.

The key points

  • 1.IAR separates knowledge injection, QA alignment, and recovery.
  • 2.Tests covered Common Corpus, CCI, and four model families.
  • 3.IAR beat Vanilla SFT across all metrics in 7 of 8 settings.

A paper highlighted by HF Daily Papers introduces Inject, Align, Recover, a three-stage post-training framework for retrieval-free question answering over bounded document collections. The method separates document knowledge injection, answer-only QA alignment, and recovery of general capabilities through merging with the base instruction model. Across Common Corpus and CCI tests using Llama, Phi, Qwen, and SmolLM model families, IAR improved the domain-primary and domain-general tradeoff versus conventional approaches, including Vanilla SFT in 7 of 8 dataset-model settings on all four reported metrics.

Try this today

Read the paper before replacing RAG: evaluate IAR-style post-training only on fixed corpora where retrieval-free QA is a requirement.

Sources & original reporting

This brief summarizes and links to reporting from the publishers below.

Enjoyed this brief? Get the next one in your inbox.

More in Research