Efficient Full-Context Retrieval for Long Documents
Keywords: Long-context retrieval, Efficient language models, Mamba architecture, Synthetic data generation
TL;DR: We propose an efficient retrieval model using Mamba architecture for full-context understanding of long documents.
Abstract: Long document question answering is challenging due to the quadratic computational cost of transformer-based LLMs. In resource-constrained environments, Retrieval-Augmented Generation (RAG) uses document chunking to maintain linear computational costs but often loses sight of the global context. We introduce the Mamba retriever 130M and Mamba retriever 1.3B retrievers, capable of processing entire long documents in linear time and integrating earlier context to retrieve relevant sentences for answering questions. Mamba retrievers outperform state-of-the-art embedding models across 41 long-document Q\&A benchmarks while maintaining speed and computational efficiency. Their performance is comparable to GPT-4o on long documents over 256k tokens while using significantly fewer tokens. Mamba retrievers are trained on synthetic data generated from our novel link-based method, which enhances the retrievers' ability to leverage long-range document connections. We further demonstrate the effectiveness of our link-based method over baseline synthetic data methods. All code, datasets, and model checkpoints are available at https://github.com/MambaRetriever/MambaRetriever
Primary Area: applications to computer vision, audio, language, and other modalities
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: 13829
Loading