// TOPIC STREAM
Mamba and selective state-space models.
This article explains why Mamba targets both long-sequence efficiency and language-modeling performance through selective SSMs that change state-space parameters based on the input and hardware-aware parallel algorithms.
Following Tri Dao and Albert Gu's Mamba-2 paper, this article explains the connection between state space models and attention, the hardware implications of the SSD algorithm, and the claims that remain unverified.
Mamba-3 seeks to improve both efficiency and state-tracking ability from a state-space perspective through better recurrence, complex-valued state updates, and multi-input/multi-output I/O. This article analyzes the design and evidence boundary, focusing on the 1.5B results stated in the paper's abstract.
Hidden State Poisoning Attacks against Mamba-based Language Models analyzes how a short trigger can overwrite the hidden state of Mamba-family models and break information retrieval. Centered on HiSPA and RoBench-25, this article explains the relationship between efficient state and the security boundary within the scope of the original abstract.