An AI training method that separates structured document injection, answer-only question-and-answer alignment and post-hoc recovery through model merging scored higher than direct fine-tuning on all four reported measures in seven of eight main dataset-model settings, according to an arXiv preprint. Across those settings, average domain-question accuracy was 3.6 percentage points higher and mean general performance was 12.1 points higher.
The study’s retrieval-free setting asks models to answer domain questions from knowledge internalized from fixed document collections, without retrieving documents for each question. IAR—short for Inject, Align, Recover—keeps those three training jobs separate.
The comparison
The main comparison covered two document collections, labeled CC and CCI. CC had 14,258 training QA pairs and 750 test pairs; CCI had 10,926 training pairs and 575 test pairs. Both were run on Llama-3.2-3B, Phi-4-mini, Qwen3-4B and SmolLM3-3B, while Qwen3-8B, Qwen3-14B and Qwen3-32B appeared in a separate CC scaling analysis.
Vanilla SFT was the main comparator. The wider baseline set included the original instruction model, BudgetMatch, SDFT, LoRA, Replay, Base-initialized CPT+SFT and FAPM.
For the recovery step, the authors fixed a 12-candidate grid covering SLERP, task arithmetic, TIES and DARE settings, selected an operating point using validation data, and then reported it on held-out test data.
Where the scores moved
On CC, IAR scored higher than Vanilla SFT on domain accuracy and each of the three general metrics for all four model families. On CCI, that pattern held for Llama, Qwen3-4B and SmolLM3-3B.
One concrete comparison came from CC’s Qwen3-4B run: domain accuracy was 50.5% with IAR versus 42.4% with Vanilla SFT, while IAR also scored higher on all three general metrics.
Against BudgetMatch, IAR had higher domain accuracy in three of four settings, higher mean general performance in all four, and won 14 of 16 metric comparisons. The CC Llama result remained a trade-off between domain and general performance.
Recovery changes the balance
Domain gains were already present before recovery. The best Inject+Align checkpoints were higher than Vanilla SFT by 2.8, 7.7, 5.3 and 4.7 percentage points on CC for Llama, Phi, Qwen3-4B and SmolLM, respectively; on CCI, the corresponding gains were 5.6, 6.1, 0.4 and 2.3 points.
In the separate Qwen3 scaling analysis on CC, the selected recovery checkpoint stayed within 1.1 percentage points of Best IA on domain accuracy across Qwen3-8B, Qwen3-14B and Qwen3-32B, while mean general performance was 14.9 to 24.1 points higher. The authors describe IAR as a strong, setting-dependent operating point rather than a universally dominant recipe, and limit the scaling result to this repeated CC pattern.
Why the finding remains provisional
The evaluation’s judging audit showed imperfect agreement. Across 242,255 evaluated model-answer instances, the first two judges agreed exactly in 70.7% of cases; after scores were collapsed into binary outcomes, agreement was 84.8%, Cohen’s kappa was .691, and a third judge was triggered in 29.7% of cases.
The authors treat those LLM judgments as an auditable instrument rather than a gold-label source. Every reported checkpoint also came from a single training run, with no averaging over multiple seeds, and the report gives no confidence interval or p-value for the average score differences.
The Qwen scaling result carries an additional provenance caveat: the available run archive lacks the original Inject and Align training arguments, logs and node manifests, so several training settings were not inferred from launcher defaults.
The document is an arXiv version 1 preprint dated 20 Aug 2026. Its evidence is limited to offline retrieval-free QA and general-benchmark evaluations on CC and CCI checkpoints; it does not test retrieval-based deployment, other corpora or independent external evaluation.
The findings compare the tested setups; they do not establish that IAR caused the gains or that the recipe will dominate elsewhere. Whether the pattern survives repeated runs, additional corpora and other evaluation settings remains open.
Paper data and sources
Original title: Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization
Authors: Qian Kou, Xiaofeng Shi, Xiaosong Qiu, Hua Zhou
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text