文章背景与核心概要
检索增强生成（RAG）被广泛用于将特定语料库的知识引入到指令遵循语言模型（Instruct LLM）中，但当检索失败时，它往往会导致模型产生幻觉。传统的参数注入方法（如扩展预训练 EPT）虽然可以注入新知识，但会损害 Instruct LLM 的指令遵循能力——这迫使人们随后必须进行昂贵且往往难以实现的指令微调（IFT）。
为了克服这些局限性，作者推出了 DKL（Decoupled Knowledge Learning，解耦知识学习）。DKL 不直接对 Instruct LLM 进行扩展预训练，而是对其对应的基础大语言模型（Base LLM）进行 EPT。随后，通过模型合并（Model Merging）技术将注入了知识的权重合并回 Instruct LLM 中。这种轻量级方法成功注入了特定语料库的知识，同时完美保留了模型的指令遵循能力，在检索失败的情况下显著提升了 RAG 的准确率，并且比以往方法消耗了少得多的训练数据。
DKL: Decoupled Knowledge Learning for Instruction-Tuned Language Models
Summary
Retrieval-Augmented Generation (RAG) is widely used to ground instruction-following language models (Instruct LLM) in specific corpora, but it often leads to hallucinations when retrieval fails. Traditional parameter-injection methods like Extended Pre-Training (EPT) can infuse new knowledge, but they compromise an Instruct LLM's instruction-following capabilities—forcing expensive and often infeasible instruction fine-tuning (IFT) afterwards.
To overcome these limitations, the authors introduce DKL (Decoupled Knowledge Learning). DKL performs EPT on the model's corresponding base LLM rather than the Instruct LLM. The resulting knowledge-infused weights are then merged back into the Instruct LLM via model merging. This lightweight approach successfully infuses corpus-specific knowledge while fully preserving the model's instruction-following capabilities, significantly boosting RAG accuracy on retrieval failure cases while using substantially less training data than prior methods.
检索增强生成（RAG）被广泛用于将特定语料库的知识引入到指令遵循语言模型（Instruct LLM）中，但当检索失败时，它往往会导致模型产生幻觉。传统的参数注入方法（如扩展预训练 EPT）虽然可以注入新知识，但会损害 Instruct LLM 的指令遵循能力——这迫使人们随后必须进行昂贵且往往难以实现的指令微调（IFT）。
为了克服这些局限性，作者推出了 DKL（Decoupled Knowledge Learning）。DKL 不直接对 Instruct LLM 进行扩展预训练，而是对其对应的基础大语言模型（Base LLM）进行 EPT。随后，通过模型合并技术将注入了知识的权重合并回 Instruct LLM 中。这种轻量级方法成功注入了特定语料库的知识，同时完美保留了模型的指令遵循能力，在检索失败的情况下显著提升了 RAG 的准确率，并且比以往方法消耗了少得多的训练数据。
Paper Metadata
- Title: DKL: Decoupled Knowledge Learning for Instruction-Tuned Language Models
- Authors: Kushagra Bhushan, Meghanadh Pulivarthi, Sai Krishna Reddy Sathi, Gaurav Pandey, Sonam Gupta, Vineet Kumar, Jaydeep Sen, Yatin Nandwani, Sachindra Joshi, Dinesh Raghu
- Submitted: September 2, 2026
- Primary Subjects: Computation and Language (
cs.CL); Artificial Intelligence (cs.AI) - Identifiers: arXiv:2609.02685 [cs.CL] | DOI: 10.48550/arXiv.2609.02685
- Stats: 20 pages, 4 figures, 15 tables
Abstract
RAG has become the de facto method for incorporating new, corpus-specific knowledge into an instruction following LLM (Instruct LLM). Although RAG-based prompting improves factual grounding, it fails when retrieval is incorrect or incomplete, leading to hallucinations. Finetuning methods such as RAFT and PA-RAG enhance RAG by injecting new knowledge into the model's parameters, but require generating a massive amount of synthetic QA that covers the entire corpus. Extended Pre-Training (EPT) on the text corpus avoids the need for comprehensive synthetic data generation but compromises an Instruct LLM's instruction-following capabilities, necessitating instruction fine-tuning (IFT) after pre-training. However, IFT is costly and may be infeasible due to the unavailability of an instruction-tuning corpus.
In this work, we propose DKL—Decoupled Knowledge Learning for Instruction-Tuned Language Models. Instead of doing EPT on the Instruct LLM, DKL performs EPT on its corresponding base LLM to infuse new knowledge. These knowledge infused weights are then merged with the Instruct LLM, imparting new knowledge without affecting their instruction-following capabilities. DKL is a lightweight method that avoids expensive instruction fine-tuning and relies on model merging to infuse the new knowledge into the Instruct LLM without destroying its instruction following capabilities. Empirical results show that DKL improves RAG accuracy from 54.17 to 79.26 on retrieval failure cases, while outperforming prior approaches with substantially less training data.
RAG 已成为将新的、特定语料库的知识并入指令遵循 LLM（Instruct LLM）的事实标准方法。尽管基于 RAG 的提示词增强了事实依据，但当检索错误或不完整时它便会失效，从而导致幻觉。微调方法（如 RAFT 和 PA-RAG）通过将新知识注入模型参数来增强 RAG，但需要生成覆盖整个语料库的大量合成问答对。在文本语料库上进行扩展预训练（EPT）避免了对全面合成数据生成的需要，但它会损害 Instruct LLM 的指令遵循能力，从而在预训练后必须进行指令微调（IFT）。然而，IFT 成本高昂，且由于缺乏指令微调语料库，可能根本无法实现。
在这项工作中，我们提出了 DKL——面向指令微调语言模型的解耦知识学习方法（Decoupled Knowledge Learning for Instruction-Tuned Language Models）。DKL 没有对 Instruct LLM 进行 EPT，而是对其对应的基础 LLM（base LLM）进行 EPT 以注入新知识。接着，将这些注入了知识的权重与 Instruct LLM 进行合并，在不影响其指令遵循能力的前提下赋予其新知识。DKL 是一种轻量级方法，它避免了高昂的指令微调开销，并依靠模型合并将新知识注入到 Instruct LLM 中，同时不会破坏其指令遵循能力。实验结果表明，在检索失败的情况下，DKL 将 RAG 的准确率从 54.17 提升至 79.26，同时以显著较少的训练数据优于以往的方法。
Access the Paper
- PDF: View PDF
- Source: TeX Source
- HTML View: arXiv HTML (Experimental)
