The paper "MambaOut: Do We Really Need Mamba for Vision?" by Weihao Yu and Xinchao Wang from the National University of Singapore explores the effectiveness of the Mamba architecture, which incorporates a state space model (SSM) inspired by RNN-like token mixers, in vision tasks. The authors argue that while Mamba has shown promise in addressing the quadratic complexity of attention mechanisms, its performance in vision tasks, particularly image classification, is often underwhelming compared to convolutional and attention-based models. They hypothesize that Mamba is not necessary for image classification tasks, which do not align with the long-sequence and autoregressive characteristics that Mamba is designed to handle. To test this hypothesis, they construct a series of models named *MambaOut* by stacking Mamba blocks while removing their core token mixer, SSM. Experimental results show that MambaOut outperforms visual Mamba models on ImageNet image classification, confirming that Mamba is indeed unnecessary for this task. However, MambaOut does not match the performance of state-of-the-art visual Mamba models in detection and segmentation tasks, highlighting the potential of Mamba for long-sequence visual tasks. The paper concludes that MambaOut can serve as a natural baseline for future research on visual Mamba models.The paper "MambaOut: Do We Really Need Mamba for Vision?" by Weihao Yu and Xinchao Wang from the National University of Singapore explores the effectiveness of the Mamba architecture, which incorporates a state space model (SSM) inspired by RNN-like token mixers, in vision tasks. The authors argue that while Mamba has shown promise in addressing the quadratic complexity of attention mechanisms, its performance in vision tasks, particularly image classification, is often underwhelming compared to convolutional and attention-based models. They hypothesize that Mamba is not necessary for image classification tasks, which do not align with the long-sequence and autoregressive characteristics that Mamba is designed to handle. To test this hypothesis, they construct a series of models named *MambaOut* by stacking Mamba blocks while removing their core token mixer, SSM. Experimental results show that MambaOut outperforms visual Mamba models on ImageNet image classification, confirming that Mamba is indeed unnecessary for this task. However, MambaOut does not match the performance of state-of-the-art visual Mamba models in detection and segmentation tasks, highlighting the potential of Mamba for long-sequence visual tasks. The paper concludes that MambaOut can serve as a natural baseline for future research on visual Mamba models.