注入、对齐、恢复：面向无检索文档知识内化的分阶段后训练

摘要

大型语言模型在推理阶段不检索源文档时，往往无法回答关于限定范围文档集合的问题。我们将这一场景研究为文档知识内化：将固定语料库转化为可用的参数化知识，以实现无检索问答。为此，我们提出 IAR（Inject, Align, and Recover，注入、对齐与恢复），一种三阶段后训练框架，将结构化文档知识注入、问答行为对齐和通用能力恢复分离开来。与常规持续预训练不同，Inject 阶段将源文档转换为续写、改写和指令条件重建目标；Align 阶段随后使用仅答案的问答监督适配注入后的模型；Recover 阶段则将领域适配模型与基础指令模型合并，以恢复通用能力。在 Common Corpus (CC) 和 CCI 上，以及 Llama、Phi、Qwen 和 SmolLM 模型族中，IAR 改善了无检索文档内化的领域主要能力与领域通用能力前沿。在主要对比中，IAR 在 8 个数据集-模型组合中的 7 个上，所有四项报告指标均优于普通 SFT；领域问答准确率平均提升 3.6 个百分点，在 IFEval、MMLU 和 MSBench 上的平均通用性能提升 12.1 个百分点。扩展的 CC 基线显示，LoRA 和 FAPM 可以在个别通用指标上获胜，但在同时达到领先或接近领先领域内化水平的方法中，IAR 仍然保持着最强的通用能力表现之一。

English

Large language models often fail to answer questions about a bounded document collection when the source documents are not retrieved at inference time. We study this setting as document knowledge internalization: converting a fixed corpus into usable parametric knowledge for retrieval-free question answering. We propose IAR (Inject, Align, and Recover), a three-stage post-training framework that separates structured document knowledge injection, QA behavior alignment, and general ability recovery. Unlike conventional continued pretraining, Inject converts source documents into continuation, rewrite, and instruction-conditioned reconstruction objectives. Align then adapts the injected model with answer-only QA supervision, while Recover merges the domain-adapted model with the base instruction model to recover general capabilities. Across Common Corpus (CC) and CCI, and across Llama, Phi, Qwen, and SmolLM model families, IAR improves the domain-primary domain-general frontier for retrieval-free document internalization. In the main comparison, 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. Extended CC baselines show that LoRA and FAPM can win individual general metrics, but among methods that also reach leading or near-leading domain internalization, IAR retains one of the strongest general profiles.