Memba: Membrane-Driven Parameter-Efficient Fine-Tuning for Mamba

Published at ICLR 2026 🎉 [Paper]

Overview

State Space Models (SSMs) like Mamba have emerged as powerful alternatives to Transformers, but existing PEFT methods simply transfer Transformer-tailored approaches without addressing Mamba's unique temporal processing dynamics.

Memba introduces a membrane-driven PEFT approach specifically designed for Mamba, with three key components:

1. Leaky Integrate Membrane (LIM) — Bio-inspired gating mechanism that accumulates membrane potentials over time on Mamba's gate branch, enhancing selective information retention
2. Strategic LoRA Placement — Low-Rank Adaptations on input and output projections
3. Cross-Layer Membrane Transfer — Averaged membrane states propagated across layers for temporal coherence

Results

Language Tasks (Commonsense Reasoning)

Model	Method	#Params(%)	Avg. Acc. (%)
Mamba-370M	LoRAp (X)	2.67	47.8
Mamba-370M	Memba (in_proj)	2.07	48.2
Mamba-790M	LoRAp (X)	1.75	50.8
Mamba-790M	Memba (in_proj)	1.47	51.8
Mamba-1.4B	LoRAp (X)	1.36	53.7
Mamba-1.4B	Memba (in_proj)	1.13	54.5

Vision Tasks (VTAB-1k with Vim-S)

Method	#Params(K)	Natural	Specialized	Structured	Avg.
LoRAp (X)	1,778	76.64	83.89	60.84	71.52
LoRA (out_proj)	2,663	76.42	83.96	60.08	71.12
Hybrid (w/ proj)	117,236	77.00	84.41	61.55	72.05
Memba (in+out proj)	4,718	77.07	85.66	61.70	72.40

Getting Started

For setup, dependencies, and detailed usage instructions, please refer to the README in each task directory:

• Language Tasks: language/commonsense_reasoning/README.md
• Vision Tasks: vision/README.md

Project Structure

Memba/
├── language/
│   └── commonsense_reasoning/
│       ├── finetune.py                 # Language fine-tuning entry point
│       ├── lm_harness_eval.py          # Evaluation with lm-evaluation-harness
│       ├── requirements.txt
│       └── mamba_peft/src/peft/
│           └── tuners/mamba_peft.py    # Memba PEFT implementation (LIM + LoRA)
├── vision/
│   ├── tools/                          # Training and testing scripts
│   ├── configs/mmpretrain/vim/vtab1k/  # Experiment configs
│   └── personal_lib/
│       ├── mmpretrain/models/peft/     # Vision Memba implementation
│       └── external_packages/mamba-1p1p1/  # Modified Mamba with LIM integration

Citation

@article{lee2025memba,
  title={Memba: Membrane-driven Parameter-Efficient Fine-Tuning for Mamba},
  author={Lee, Donghyun and Li, Yuhang and Yin, Ruokai and Xiao, Shiting and Panda, Priyadarshini},
  journal={arXiv preprint arXiv:2506.18184},
  year={2025}
}

Acknowledgements

Our code builds upon Mamba, MambaPEFT, Vim, MMPreTrain, LLM-Adapters, and PEFT. Especially, we built Memba upon MambaPEFT code. Thank you for the authors!

License

This project is released under the Apache 2.0 License.