Mamba

Mamba studies sequence models that update a compact recurrent state selectively, aiming to capture long-range context with less memory and computation than attention. Work explores bidirectional and Mamba-attention hybrids, scan orders and state-space variants, as well as combinations with convolutional networks for images, audio, language, and time series. These approaches are evaluated on settings such as long-context language tasks, speech recognition, time-series benchmarks, and medical-image segmentation. A central tension is whether efficient state propagation preserves fine local details and robust long-range reasoning; performance can depend on sequence ordering, task length, and when attention or local feature processing is added.