Research map / Architectures and efficiency / Pruning and efficient tuning
Memba: Membrane-driven Parameter-Efficient Fine-Tuning for Mamba
Donghyun Lee, Yuhang Li, Ruokai Yin, Shiting Xiao, Priyadarshini Panda
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 LIM neurons and membrane transfer on top of LoRA (ablation) | Memba · Commonsense reasoning (Mamba 130M) | accuracy: 43.8 -> 44.3 | Improved |
| add LIM neurons and membrane transfer on top of LoRA (ablation) | Memba · VTAB-1k Specialized (Vim-S) | accuracy: 85.3 -> 85.7 | Improved |
| add cross-layer membrane transfer to LIM + LoRA (ablation) | Memba · Commonsense reasoning (Mamba 130M) | accuracy: 44.0 -> 44.3 | Improved |
| add LoRA to LIM + membrane transfer (ablation) | Memba · VTAB-1k Specialized (Vim-S) | accuracy: 84.0 -> 85.7 | Improved |
| apply LoRA to both in_proj and out_proj instead of one (ablation) | Memba · Commonsense reasoning (Mamba 790M) | accuracy: 52.1 -> 52.3 | Improved |
| apply both in_proj and out_proj LoRA instead of in_proj only (ablation) | Memba · VTAB-1k (Vanilla-VMamba-S) | accuracy: 73.64 -> 73.48 | Hurt |
| exclude in_proj or out_proj from LoRA placement (ablation) | Memba-130M · Commonsense reasoning | accuracy: -1.2% (in), -0.8% (out) | Hurt |
| raise the leaky factor tau from 1/5 to 1/2 (ablation) | LIM neuron · VTAB-1k Specialized (Vim-S) | accuracy: 85.0 -> 85.7 | Improved |
| set threshold V_th to 1 rather than 0.5, 2 or 3 (ablation) | LIM neuron · VTAB-1k Specialized (Vim-S) | accuracy: 85.4 / 85.3 / 85.2 -> 85.7 | Improved |
| use Memba (in+out_proj) instead of full fine-tuning | Mamba 790M · Commonsense reasoning | accuracy: 51.6 -> 52.3 | Improved |
| use Memba (in+out_proj) instead of full fine-tuning | Vim-S · VTAB-1k (avg) | accuracy: 50.08 -> 72.40 | Improved |
| use Memba (in+out_proj) instead of best prior PEFT Hybrid | Vim-S · VTAB-1k (avg) | accuracy: 72.05 -> 72.40 | Improved |
Closest papers by meaning
- Parameter Efficient Mamba Tuning via Projector-targeted Diagonal-centric Linear Transformation (CVPR 2025 · similarity 0.88)
- Understanding Input Selectivity in Mamba: Impact on Approximation Power, Memorization, and Associative Recall Capacity (ICML 2025 · similarity 0.87)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.85)
- Achilles' Heel of Mamba: Essential difficulties of the Mamba architecture demonstrated by synthetic data (NeurIPS 2025 · similarity 0.85)
- Coupled Mamba: Enhanced Multimodal Fusion with Coupled State Space Model (NeurIPS 2024 · similarity 0.85)
- Mamba Modulation: On the Length Generalization of Mamba Models (NeurIPS 2025 · similarity 0.85)
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.
