arXiv 2509.039342025-09-04cs.CLcs.AI

SelfAug：通过分布自对齐缓解检索增强生成中的灾难性遗忘

SelfAug: Mitigating Catastrophic Forgetting in Retrieval-Augmented Generation via Distribution Self-Alignment

University of Science and Technology of China（中国科学技术大学）；Xiaohongshu Inc.（小红书公司）

Yuqing Huang, Rongyang Zhang, Qimeng Wang, Chengqiang Lu, Yan Gao, Yi Wu, Yao Hu, Xuyang Zhi, Guiquan Liu, Xin Li, Hao Wang, Enhong Chen

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

针对检索增强生成（RAG）场景下监督微调易引发大语言模型灾难性遗忘的问题，提出自分布对齐方法SelfAug，通过对齐输入序列logits保留语义分布，实现下游学习与通用能力保留的更优平衡，并揭示分布偏移与遗忘程度的关联。

摘要：

大语言模型（LLMs）近期的进展凭借其理解和执行多样任务的卓越能力，彻底革新了自然语言处理领域。尽管监督微调，尤其是在检索增强生成（RAG）场景下的监督微调，能有效提升特定任务的性能，但它往往会导致灾难性遗忘，即模型丢失先前习得的知识和通用能力。现有解决方案要么需要访问通用指令数据，要么在保留模型原始分布方面存在局限。为克服这些局限，我们提出SelfAug——一种自分布对齐方法，该方法通过对齐输入序列的logits来保留模型的语义分布，从而缓解灾难性遗忘并提升下游性能。大量实验表明，SelfAug在下游学习与通用能力保留之间实现了更优的平衡。我们全面的实证分析揭示了RAG场景下分布偏移与灾难性遗忘严重程度之间的直接关联，凸显了通用指令调优中缺乏RAG能力如何导致微调过程中出现显著的分布偏移。我们的研究发现不仅深化了对RAG场景下灾难性遗忘的理解，还提供了一种适用于多种微调场景的实用解决方案。我们的代码已公开发布于https://github.com/USTC-StarTeam/SelfAug。

英文摘要：

Recent advancements in large language models (LLMs) have revolutionized natural language processing through their remarkable capabilities in understanding and executing diverse tasks. While supervised fine-tuning, particularly in Retrieval-Augmented Generation (RAG) scenarios, effectively enhances task-specific performance, it often leads to catastrophic forgetting, where models lose their previously acquired knowledge and general capabilities. Existing solutions either require access to general instruction data or face limitations in preserving the model's original distribution. To overcome these limitations, we propose SelfAug, a self-distribution alignment method that aligns input sequence logits to preserve the model's semantic distribution, thereby mitigating catastrophic forgetting and improving downstream performance. Extensive experiments demonstrate that SelfAug achieves a superior balance between downstream learning and general capability retention. Our comprehensive empirical analysis reveals a direct correlation between distribution shifts and the severity of catastrophic forgetting in RAG scenarios, highlighting how the absence of RAG capabilities in general instruction tuning leads to significant distribution shifts during fine-tuning. Our findings not only advance the understanding of catastrophic forgetting in RAG contexts but also provide a practical solution applicable across diverse fine-tuning scenarios. Our code is publicly available at https://github.com/USTC-StarTeam/SelfAug.

机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。