ML-Mamba: Efficient Multi-Modal Large Language Model Utilizing Mamba-2
원문 초록
Multimodal Large Language Models (MLLMs) have attracted much attention for their multifunctionality. However, traditional Transformer architectures incur significant overhead due to their secondary computational complexity. To address this issue, we introduce ML-Mamba, a multimodal language model, which utilizes the latest and efficient Mamba-2 model for inference. Mamba-2 is known for its linear scalability and fast processing of long sequences. We replace the Transformer-based backbone with a pre-trained Mamba-2 model and explore methods for integrating 2D visual selective scanning mechanisms into multimodal learning while also trying various visual encoders and Mamba-2 model variants. Our extensive experiments in various multimodal benchmark tests demonstrate the competitive performance of ML-Mamba and highlight the potential of state space models in multimodal tasks. The experimental results show that: (1) we empirically explore how to effectively apply the 2D vision selective scan mechanism for multimodal learning. We propose a novel multimodal connector called the Mamba-2 Scan Connector (MSC), which enhances representational capabilities. (2) ML-Mamba achieves performance comparable to state-of-the-art methods such as TinyLaVA and MobileVLM v2 through its linear sequential modeling while faster inference speed; (3) Compared to multimodal models utilizing Mamba-1, the Mamba-2-based ML-Mamba exhibits superior inference performance and effectiveness.
함께 살펴볼 논문
공통 태그를 근거로 연결했습니다. 실제 인용 관계나 논문의 우수성을 뜻하지 않습니다.
Sequential-Parallel Duality in Prefix Scannable Models
Specification-based Security Evaluation of Real-World Agents Tool-Calling under Long-Horizon Context Tasks
Assay2Mol: large language model-based drug design using BioAssay context
Open Source Planning & Control System with Language Agents for Autonomous Scientific Discovery
Differential Mamba
Detecting Referring Expressions in Visually Grounded Dialogue with Autoregressive Language Models
토론
이 논문과 연결된 포럼 글이 아직 없습니다.