LongMamba: Enhancing Mamba's Long-Context Capabilities via Training-Free Receptive Field Enlargement
Keywords: Large Language Model, Long Context Understanding
Abstract: Mamba models have emerged as an efficient alternative to Transformer models for language modeling tasks, offering linear complexity as context length increases. However, despite their efficiency in handling long contexts, recent studies have demonstrated that Mamba models underperform in understanding extended contexts compared to Transformer models. To address this significant shortfall, we propose ``LongMamba", a training-free technique that significantly enhances the long-context capabilities of Mamba models. Our approach builds upon the discovery that hidden state channels in Mamba models—categorized into \textit{local} and \textit{global channels} based on their receptive field lengths—exhibit distinct functionalities. Specifically, the \textit{global channels} struggle to adaptively extend their effective receptive fields when input lengths far exceed their training sequence length due to exponential decay in their hidden states. We hypothesize this exponential decay is the root cause of Mamba models’ limited performance in extended contexts. LongMamba counters this by effectively expanding the \textit{global channels}' receptive fields to fully encompass the input sequence length, thus enabling them to capture global information more effectively. Through extensive benchmarking across synthetic and real-world long-context scenarios, LongMamba sets a new standard for state-of-the-art performance in Mamba-based long-context tasks, significantly extending the operational range of Mamba models without requiring additional fine-tuning. All code and models will be released upon acceptance.
Primary Area: foundation or frontier models, including LLMs
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: 13423
Loading