Retrieval-Augmented Editing Generation: Impact of Knowledge Editing and Fine-Tuning on RAG
Keywords: Retrieval-Augmented Generation, Knowledge Editing, Parameter-Efficient Fine-Tuning
Abstract: The knowledge embedded in Large Language Models (LLMs) is static, tied to the time when the training data was collected.
While Retrieval-Augmented Generation (RAG) methods are widely used to introduce new knowledge, they simply rely on retrieved information for reasoning without integrating it into the model’s parameters. This limits the model's ability for long-term knowledge retention and autonomous learning.
To overcome this, in this work, we propose the \textbf{R}etrieval-\textbf{A}ugmented \textbf{E}diting \textbf{G}eneration (RAEG) framework for open-domain question answering (ODQA) tasks.
RAEG enhances model generation performance by first editing the retrieved paragraphs to inject necessary knowledge, followed by an augmented generation phase. This dual mechanism—combining knowledge injection and retrieval augmentation—provides complementary advantages in the reasoning process. When the injected knowledge alone is insufficient for accurate generation, the model can rely on the retrieved information to compensate, and conversely, when retrieval yields suboptimal results, the injected knowledge ensures continuity and accuracy in the response. This interplay between internalized and externally sourced knowledge reinforces the model's ability to produce correct answers, thereby enhancing overall task performance.
We explore the impact of two key methods for knowledge injection: Knowledge Editing (KE) and Parameter-Efficient Fine-Tuning (PEFT), and analyze how modifying the model's parameters influences its reasoning abilities and generation outcomes. To further improve RAEG's performance, we introduce a re-ranking mechanism to optimize the integration of external knowledge and apply parameter pruning to mitigate the potential drawbacks of parameter modifications during KE.
Evaluations on two authoritative ODQA benchmarks show that RAEG is able to further replace RAG as a competitive method.
Our data and code will be available at \url{https://github.com/XXX/XXX}.
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 13471
Loading