2Mamba2Furious: Linear in Complexity, Competitive in Accuracy
Abstract
Researchers enhance linear attention by simplifying Mamba-2 and improving its architectural components to achieve near-softmax accuracy while maintaining memory efficiency for long sequences.
Linear attention transformers have become a strong alternative to softmax attention due to their efficiency. However, linear attention tends to be less expressive and results in reduced accuracy compared to softmax attention. To bridge the accuracy gap between softmax attention and linear attention, we manipulate Mamba-2, a very strong linear attention variant. We first simplify Mamba-2 down to its most fundamental and important components, evaluating which specific choices make it most accurate. From this simplified Mamba variant (Mamba-2S), we improve the A-mask and increase the order of the hidden state, resulting in a method, which we call 2Mamba, that is nearly as accurate as softmax attention, yet much more memory efficient for long context lengths. We also investigate elements to Mamba-2 that help surpass softmax attention accuracy. Code is provided for all our experiments
Community
Linear attention transformers have become a strong alternative to softmax attention due to their efficiency. However, linear attention tends to be less expressive and results in reduced accuracy compared to softmax attention. To bridge the accuracy gap between softmax attention and linear attention, we manipulate Mamba-2, a very strong linear attention variant. We first simplify Mamba-2 down to its most fundamental and important components, evaluating which specific choices make it most accurate. From this simplified Mamba variant (Mamba-2S), we improve the A-mask and increase the order of the hidden state, resulting in a method, which we call 2Mamba, that is nearly as accurate as softmax attention, yet much more memory efficient for long context lengths. We also investigate elements to Mamba-2 that help surpass softmax attention accuracy. Code is provided for all our experiments
Wait, if we're able to get softmax-like performance on NIAH problems, isn't that a groundbreaking result? Since the entire problem we have with SSMs was their inability to remember 'facts'.
I do think it's a cool result, but I don't think any massive conclusions should be made until context usage is empirically evaluated across a sweep of multiple model sizes, data mixtures, and context lengths.
This is an automated message from the Librarian Bot. I found the following papers similar to this paper.
The following papers were recommended by the Semantic Scholar API
- Neural Attention Search Linear: Towards Adaptive Token-Level Hybrid Attention Models (2026)
- ZeroS: Zero-Sum Linear Attention for Efficient Transformers (2026)
- Online Vector Quantized Attention (2026)
- Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contexts (2026)
- Softmax Linear Attention: Reclaiming Global Competition (2026)
- LUCID: Attention with Preconditioned Representations (2026)
- MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head (2026)
Please give a thumbs up to this comment if you found it helpful!
If you want recommendations for any Paper on Hugging Face checkout this Space
You can directly ask Librarian Bot for paper recommendations by tagging it in a comment:
@librarian-bot
recommend
Models citing this paper 0
No model linking this paper
Datasets citing this paper 0
No dataset linking this paper
Spaces citing this paper 0
No Space linking this paper