注入、對齊、恢復：分階段後訓練實現免檢索文件知識內化

摘要

大型語言模型在推論時若未檢索來源文檔，往往無法回答關於特定文件集的問題。我們將此情境研究為文檔知識內化：將固定語料庫轉換為可用的參數化知識，以實現無檢索問答。我們提出 IAR（Inject, Align, and Recover，注入、對齊與恢復），這是一個三階段的後訓練框架，將結構化文檔知識注入、問答行為對齊與一般能力恢復分離開來。與傳統的持續預訓練不同，Inject 階段將來源文檔轉換為續寫、改寫與指令條件式重建目標。Align 階段隨後以僅含答案的問答監督方式調整注入後的模型，而 Recover 階段則將領域適應後的模型與基礎指令模型合併，以恢復一般能力。在 Common Corpus（CC）與 CCI 上，以及 Llama、Phi、Qwen 和 SmolLM 模型家族中，IAR 皆改善了無檢索文檔內化的領域主要—領域一般能力前緣。在主要比較中，IAR 在 8 個資料集—模型設定中的 7 個，於全部四項報告指標上優於 Vanilla SFT，領域問答準確率平均提升 3.6 個百分點，而 IFEval、MMLU 與 MSBench 的一般能力平均值則提升 12.1 個百分點。擴展的 CC 基線結果顯示，LoRA 與 FAPM 可在個別一般能力指標上勝出，但在同時達到領先或接近領先領域內化表現的方法中，IAR 仍保有最強的一般能力輪廓之一。

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.