State-space model
A family of sequence models that carry a fixed-size running state instead of attending to every past token.
Also: SSM, Mamba
State-space models, borrowed from control theory, process a sequence by updating a compact internal state at each step, so memory and compute grow only in line with length. Albert Gu and colleagues made them competitive for language with S4 in 2021, and Gu and Tri Dao's Mamba in 2023 added selection, letting the state decide what to keep. They are fast on very long texts but have struggled with tasks that need exact recall of earlier passages, so many recent models combine them with attention layers.