Selective SSMs

Selective state-space models study how recurrent systems can use each input to decide what information to write, retain, or forget, making long sequences cheaper to process than with full self-attention. Mamba and related selective-scan architectures are applied to language, time series, vision, and multimodal tasks, often alongside convolution, attention, frequency-domain features, or graph structure. The field also develops theoretical analyses of memory, in-context learning, stability, and interpretability, and evaluates models on long-context retrieval, forecasting, and task-specific benchmarks. A central tension is that efficient, scalable recurrence can perform strongly, yet may struggle with content-based retrieval or learned memory use, and its advantages depend on task, training, and data.