Retrieval‑free internalization now beats standard fine‑tuning on domain‑specific question answering, and it does so without sacrificing the model’s broad linguistic competence. The IAR framework makes this possible by turning a static document collection into parametric knowledge that lives directly in the weights.
Before IAR, most systems relied on an external retriever at inference time, coupling a dense index with a language model or augmenting the model via LoRA adapters to expose domain data. Those approaches still incurred latency from the retrieval step and often required separate pipelines for general‑purpose tasks.
IAR improves over Vanilla SFT on all four reported metrics in 7 of 8 dataset‑model settings, with average gains of 3.6 percentage points in domain QA accuracy and 12.1 percentage points in mean general performance across IFEval, MMLU, and MSBench[1]. The three‑stage post‑training—Inject, Align, Recover—delivers a consistent uplift while keeping the model usable for unrelated benchmarks.
Qwen3‑4B CC is the clearest example, reaching 50.5 % versus 42.4 % domain accuracy while improving all three general metrics[1]. This single model jump illustrates how weight‑level knowledge injection can eclipse a vanilla supervised fine‑tune even on a mid‑size 4B parameter backbone.
Relative to Vanilla SFT, Best IA improves domain accuracy by 2.8, 7.7, 5.3, and 4.7 points on CC for Llama, Phi, Qwen3‑4B, and SmolLM respectively; the corresponding CCI gains are 5.6, 6.1, 0.4, and 2.3 points[1]. The injection‑and‑align stages alone already capture a sizable fraction of the final performance boost before any recovery step.
The paper’s own limitation is that achieving these gains requires a three‑stage training pipeline, which adds engineering overhead and may not scale trivially to corpora larger than those tested. This suggests an open question: can a single‑pass or lighter‑weight internalization procedure retain both domain fidelity and general capability at scale?
If the reported improvements hold across broader settings, developers should replace external retriever components with IAR‑enhanced models for latency‑critical services, cutting inference time and infrastructure complexity while preserving—or even improving—answer quality.
