New Mamba model variant enhances memory retention and bilinear computation

Researchers have introduced Bilinear Input Modulation (BIM) to enhance Selective State Space Models (SSMs), specifically Mamba, by incorporating state-input products. This augmentation allows for improved memory retention and multiplicative computation, addressing limitations in Mamba's diagonal state transitions. The proposed methods, including Coupled Bilinear Input Modulation (seq-BIM) and Parallel Bilinear Input Modulation (p-BIM), demonstrate significant performance gains on tasks requiring memory and bilinear processing, outperforming simpler gating mechanisms. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Introduces a new method to improve memory retention and computational capacity in state-space models.

RANK_REASON Academic paper introducing a novel computational technique for existing models.

Read on arXiv cs.LG →

• Mamba
• Koopman
• SSM
• arXiv
• Hiroki Fujii

• paper
• other