Tech Meridian ← ENTITY INDEX
PROMY MERIDIAN RU

COMPANY · ENTITY #5477

State Space Models (SSMs)

Related event timeline, sources and context from the news index.

EVENT TIMELINE

4

RESEARCH · 1 SOURCE · Apple Machine Learning Research

On the Importance of Gating: how gating steers memorization vs in‑context learning in SSMs

Researchers show theoretically and empirically that the gating mechanism in state space models (SSMs) often drives training toward an in‑weights memorization solution and can delay or prevent convergence to correct in‑context learning, while also sometimes improving generalization to longer sequences. The paper analyzes how gating shapes both training dynamics and long‑sequence generalization in linear‑time sequence models.

7.0

MODELS · 1 SOURCE · AWS Machine Learning

Cartesia demonstrates SSM-based Sonic and Ink for sub-100ms streaming voice on AWS

An AWS post describes how Cartesia, an AWS Partner, uses State Space Models (SSMs) and AWS infrastructure to run two production streaming voice models—Sonic (text-to-speech) and Ink (speech-to-text)—optimized for sub-100ms first-byte latency and global scale. Sonic is presented as a streaming-native TTS with roughly 90ms to first audio byte and multi-language/voice-cloning features; Ink is positioned as a streaming STT with native turn detection for conversational use cases.

6.0

MODELS · 1 SOURCE · Apple Machine Learning Research

Attention to Mamba: two-stage recipe distills Pythia-1B Transformer into Mamba SSM

Authors propose a principled two-stage cross-architecture distillation: first distill a Transformer into a linearized-Attention model (using a kernel-trick adaptation), then distill that into a Mamba-style State Space Model without Attention. The distilled Mamba nearly matches the Pythia-1B teacher (perplexity 14.01 vs 13.86) across evaluation and includes ablations, scaling analyses, and a distillation scaling law for compute-optimal allocation.

7.0

MODELS · 1 SOURCE · arXiv cs.AI

DSRec: dual-interest cross-SSM for sequential recommendation

arXiv:2609.21548v1 introduces DSRec, a dual-interest cross-SSM architecture that disentangles long-term and short-term item roles by encoding items into long-term interest embeddings (processed by a full-sequence Mamba SSM) and a time-modulated short-term SSM, with a residual cross-fusion mechanism to align the two branches; experiments on public benchmarks report improvements over prior state-of-the-art methods.

5.0