FM-Mamba paper proposes native 2D state space model with stability-aware step sizes
An arXiv abstract describes a Mamba variant that processes images directly on the 2D grid, with input-dependent step sizes constrained by a stability criterion.
In plain words
Mamba-style models usually turn an image into a long one-dimensional sequence before processing it, which loses some spatial structure. FM-Mamba instead runs its recurrence across both image axes at once, and it caps the model's per-position step sizes using a mathematical stability rule so signals do not blow up locally.
What was announced
The authors frame the problem as a trade-off in existing Mamba-based approaches for 2D features: flattening image feature maps into 1D sequences and fusing multiple scans breaks spatial adjacency and multiplies computation, while recursing directly on the 2D grid has relied on a fixed, input-independent rescaling factor and gathered information only from predecessors on a single anti-diagonal, leaving the discrete step unconstrained by 2D stability theory.
FM-Mamba is built on the Fornasini–Marchesini model, with a coupled recurrence that evolves along both spatial axes and two independent, input-dependent step sizes governing propagation along each axis, so coefficients adapt per position and per direction. A normalized ratio yields a closed-form piecewise bound on the two step sizes; configurations outside the stable interval are projected onto this bound, and the criterion is re-verified per position after training, which the authors say suppresses systematic local amplification without global damping. Computation uses a dual-diagonal wavefront scan in which each update gathers information from two anti-diagonals, and the input enters through a single projection, so neither training nor inference flattens the feature map. Code is hosted on GitHub.
What it does not show
- The evidence is the paper's abstract only; no benchmark results, comparisons, or independent evaluation are included in the supplied passages.
For security and IT teams
- Access
- The abstract links to a public code repository at github.com/BitSsm/FM-Mamba-Runtime.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.