// TOPIC STREAM
Papers, experiments, evidence, and technical analysis.
This article examines how the Structured State Space sequence model (S4) structures state space models to handle long dependencies and reduces them to computable Cauchy kernel operations. It distinguishes only the claims confirmed by the arXiv abstract and metadata.
Hungry Hungry Hippos decomposes the linguistic weaknesses of state space models into token recall and comparison, then explains how H3, FlashConv, and hybrid attention combinations address them.
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.
This article separates the evidence reported by the paper from the open questions of adoption, explaining how Hawk and Griffin combine recurrent-model efficiency with the Transformer's representational power.
This article explains, with evidence from the original paper, how Eagle and Finch retain RWKV-4's efficient inference path while adding matrix states, multiple heads, and data-dependent recurrence.
xLSTM redesigns LSTM gating and memory structures from the perspective of scaling modern language models. Based on the public abstract and metadata, this article examines the roles of sLSTM and mLSTM, the meaning of parallelization and memory state, and performance conditions 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.
Resona is a framework that retrieves needed information from the input context and reinjects it into a linear recurrent language model. Based on the paper's abstract, this article examines the bottleneck in context copying, the meaning of combining retrieval with recurrence, and the evaluation, cost, and evidence that must be checked before practical adoption.