Research map / Architectures and efficiency / Attention and transformers
MuonSSM: Orthogonalizing State Space Models for Sequence Modeling
Thai Khanh Nguyen, Vo Ngoc Bich Uyen, Thieu Vo, Tan Minh Nguyen, Cuong Pham
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add MuonSSM (momentum + Newton-Schulz) to Mamba | Mamba 170M · zero-shot common-sense avg (FineWeb-Edu) | avg accuracy: 42.13 -> 44.23 | Improved |
| add MuonSSM to LongHorn | LongHorn 170M · LAMBADA | perplexity: 96.80 -> 80.98 | Improved |
| add MuonSSM to LongHorn | LongHorn 170M · S-NIAH-PK 8K | accuracy: 20.0 -> 39.3 | Improved |
| add MuonSSM to Gated DeltaNet | Gated DeltaNet 170M · S-NIAH-PK 8K | accuracy: 25.7 -> 44.5 | Improved |
| add MuonSSM to LongHorn | LongHorn 170M · S-NIAH-UUID 8K | accuracy: 19.3 -> 28.6 (Mamba+Muon improves, LongHorn 2K 46.4 -> 52.9) | Mixed |
| add MuonSSM to LongHorn | LongHorn 170M · S-NIAH-PK 2K | accuracy: 67.9 -> 66.7 | Hurt |
| add MuonSSM to MambaVision-Tiny backbones | Mamba · ImageNet-1K | Top-1: 81.08 -> 81.19 | Improved |
| add MuonSSM to Mamba | MambaVision-Tiny Mamba · ImageNet-A | Top-1: 20.57 -> 20.50 | No clear effect |
| add MuonSSM to LongHorn | MambaVision-Tiny LongHorn · ADE20K | mIoU: 44.2 -> 45.7 | Improved |
| add MuonSSM to Gated DeltaNet | Gated DeltaNet · MMAct | accuracy: 66.39 -> 66.61 | No clear effect |
| add MuonSSM to Mamba | Mamba · UESTC-MMEA-CL | accuracy: 87.74 -> 91.62 | Improved |
| double state dimension of Mamba instead of using Muon (ablation) | Mamba · MMAct | accuracy: 71.47 -> 72.52 (Muon: 74.65) | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LongHorn — Longhorn: State Space Models are Amortized Online Learners (ICLR 2025) · better on 7 of 7 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection (NeurIPS 2025 · similarity 0.89)
- Elastic Spectral State Space Models for Train-Once Budgeted Inference (NeurIPS 2026 · similarity 0.88)
- MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models (NeurIPS 2026 · similarity 0.88)
- Longhorn: State Space Models are Amortized Online Learners (ICLR 2025 · similarity 0.88)
- Block-Biased Mamba for Long-Range Sequence Processing (NeurIPS 2025 · similarity 0.88)
- Bridging Expressivity and Scalability with Adaptive Unitary SSMs (NeurIPS 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
