Motif-Mamba: Enhancing Mamba with Motif-Constrained Recurrent Pathways

A new research paper on arXiv (2608.00027) introduces Motif-Mamba, a structured state space model designed to improve long-range sequence modeling. The model addresses the quadratic scaling of self-attention in large language models by augmenting Mamba's linear-time selective state space recurrence with a motif-constrained low-rank recurrent pathway. This pathway, inspired by three-node network motifs, projects hidden states into a compact dynamical subspace, applies motif-guided interactions, and maps the dynamics back to the original state space, enhancing cross-dimensional communication while preserving linear-time complexity. Experiments on long-sequence extrapolation, language modeling benchmarks, and brain-computer interface decoding show consistent improvements over baseline Mamba. The paper is authored by researchers (names not provided in the abstract) and was announced as a new submission on arXiv.