Skip to main content
    Skip to main contentSkip to navigationSkip to footer
    Artificial Intelligence

    State Space Model (SSM)

    Also known as:
    SSM
    Structured State Space
    S4-Family
    Updated: 2/11/2026

    A class of sequence models based on continuous state space theory offering linear scaling O(N) instead of quadratic attention O(N²).

    Quick Summary

    State space models model sequences as a dynamical system with O(N) scaling – the theoretical basis for Mamba and Transformer alternatives.

    Explanation

    SSMs model sequences as a linear dynamical system: x'(t) = Ax(t) + Bu(t), y(t) = Cx(t). Through discretization and special parameterization (HiPPO, S4), they can efficiently capture long dependencies. Mamba extends this with selective mechanisms.

    Marketing Relevance

    SSMs are the most promising Transformer alternative for tasks with extremely long sequences (audio, genomics, time series).

    Common Pitfalls

    Not yet fully at Transformer parity for language tasks. Less mature tooling and community. Training instabilities with naive implementation.

    Origin & History

    Gu et al. introduced HiPPO (2020) and S4 (2021). S4 first showed state-of-the-art on long-range benchmarks. Mamba (2023) made SSMs competitive for language through selective mechanisms. Mamba-2 and Jamba (2024) approached Transformer quality.

    Comparisons & Differences

    State Space Model (SSM) vs. Transformer

    Transformers use attention (O(N²), strong quality); SSMs use recurrence (O(N), more efficient for long sequences but quality gap).

    State Space Model (SSM) vs. RNN/LSTM

    RNNs have vanishing gradient; SSMs solve this through HiPPO initialization and can be trained in parallel (as convolution).

    Related Services

    Related Terms

    👋Questions? Chat with us!