{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2608.20281v1",
  "slug": "2608-20281v1-0radlbg",
  "url": "https://feed7.dev/p/2608-20281v1-0radlbg",
  "title": "Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization",
  "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.",
  "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_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.\n\nFor 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**.\n\nAverage 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.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2608.20281v1",
    "published_at": "2026-08-20T17:14:24.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "model",
  "domains": [
    "research",
    "data"
  ],
  "topics": [
    "retrieval",
    "context-engineering"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "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."
  ],
  "connected_context": {
    "meaning": "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.",
    "corpus_size": 525,
    "generated_at": "2026-08-22T21:13:03.535Z",
    "connections": [
      {
        "title": "Scaling Compute on Context — Jack Morris, Engram",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=WiqDvX6isc4",
        "feed7_url": "https://feed7.dev/p/scaling-compute-on-context-jack-morris-engram-1pxr9bt",
        "reason": "IAR operationalizes this candidate’s proposed synthetic study tasks and directly addresses its warning about overwritten prior knowledge through base-model merging."
      },
      {
        "title": "The Rise of CaaS: Context-as-a-Service for Agentic AI — Omer Primor, Bright Data",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=Ot4OPrPH4xY",
        "feed7_url": "https://feed7.dev/p/the-rise-of-caas-context-as-a-service-for-agentic-ai-omer-primor-bright-0bfgium",
        "reason": "They define complementary choices for stable repeated knowledge: owning retrieval infrastructure versus internalizing a bounded corpus into model parameters."
      },
      {
        "title": "virgiliojr94/book-to-skill",
        "source_name": "GitHub",
        "source_url": "https://github.com/virgiliojr94/book-to-skill",
        "feed7_url": "https://feed7.dev/p/book-to-skill-1av16sr",
        "reason": "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."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-08-20T17:14:24.000Z",
  "modified_at": "2026-08-20T17:14:24.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2608-20281v1-0radlbg",
    "json": "https://feed7.dev/p/2608-20281v1-0radlbg.json",
    "markdown": "https://feed7.dev/p/2608-20281v1-0radlbg.md"
  }
}