Research map / Architectures and efficiency / Attention and transformers
MaTVLM: Hybrid Mamba-Transformer for Efficient Vision-Language Modeling
Yingyue Li, Bencheng Liao, Wenyu Liu, Xinggang Wang
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 |
|---|---|---|---|
| use 25% Mamba-2 layers instead of 50% (ablation) | MaTVLM · 9-benchmark average | average score: 62.3 vs 60.1; surpassing the 50% ratio by 2.2 points | Improved |
| use 25% Mamba-2 layers instead of 12.5% (ablation) | MaTVLM · 9-benchmark average | average score: 0.5 points lower for 12.5% (61.8 vs 62.3) | Improved |
| distribute Mamba-2 layers evenly instead of all at the beginning (ablation) | MaTVLM · 9-benchmark average | average score: 62.3 vs 60.4; 1.9 points | Improved |
| distribute Mamba-2 layers evenly instead of all in the middle (ablation) | MaTVLM · 9-benchmark average | average score: 62.3 vs 58.9; 3.4 points | Improved |
| place Mamba-2 layers all at the end instead of evenly (ablation) | MaTVLM · distillation | response quality | Hurt |
| use L_prob instead of L_ce only (ablation) | MaTVLM · 9-benchmark average | average score: 55.6 -> 61.4; 5.8-point increase | Improved |
| add L_layer to L_prob (ablation) | MaTVLM · 9-benchmark average | average score: 61.4 -> 62.3; 0.9 points higher | Improved |
| add L_ce to L_prob + L_layer (ablation) | MaTVLM · 9-benchmark average | average score: 62.3 -> 61.3; 1.0-point drop | Hurt |
| distill into hybrid Mamba-2 model instead of teacher TinyLLaVA | MaTVLM · MME-P | perception score: 1466.4 -> 1484.1; 17.6-point improvement | Improved |
| replace transformer layers with Mamba-2 (25%) for inference | MaTVLM vs TinyLLaVA · generation speed | speedup: 3.6x faster | Improved |
| replace transformer layers with Mamba-2 for memory | MaTVLM vs TinyLLaVA · 16,384 tokens | peak GPU memory: 27.5% reduction | Improved |
Closest papers by meaning
- The Mamba in the Llama: Distilling and Accelerating Hybrid Models (NeurIPS 2024 · similarity 0.92)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.89)
- Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks (ICML 2024 · similarity 0.88)
- Cobra: Extending Mamba to Multi-Modal Large Language Model for Efficient Inference (AAAI 2025 · similarity 0.88)
- MambaQuant: Quantizing the Mamba Family with Variance Aligned Rotation Methods (ICLR 2025 · similarity 0.88)
- MambaVision: A Hybrid Mamba-Transformer Vision Backbone (CVPR 2025 · similarity 0.88)
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.
