GasketRAG: Systematic Alignment of Large Language Models with Retrievers
Keywords: Retrieval Augmented Generation, Large Language Models
Abstract: Retrieval-Augmented Generation (RAG) has emerged as a powerful method for enhancing the output quality of large language models (LLMs). However, existing retrievers are not specifically optimized for LLMs, and retraining them requires substantial resources. Furthermore, current approaches are often constrained to either improving the relevancy of retrieved documents or refining the documents post-retrieval. Various stages within the typical RAG pipeline present challenges in aligning LLMs with retrievers. To address these issues, we propose GasketRAG, a novel approach that introduces a gasket between the retriever and the LLM to improve their collaborative performance. By employing innovative techniques, we gather high-quality preference data and use the gasket to optimize both retrieval ranking and document refinement simultaneously. Our approach circumvents the need for constructing complex training and inference pipelines. In a fair comparison against the latest RAG methods across multiple test datasets, GasketRAG demonstrated a clear advantage.
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
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: 4036
Loading