A Unified Review of Advanced Paradigms in Spiking Neural Networks From Efficiency and Latency to Novel Architecture and Robustness

Abstract

The domain of Spiking Neural Networks (SNNs) has undergone a paradigm shift in the 2024-2025 research cycle, moving beyond theoretical validation to successfully integrate into complex domains like Large Language Models (LLMs) and advanced robotics. This review provides an exhaustive synthesis of state-of-the-art SNN optimization, focusing on Architectural Innovation, Training Algorithm Refinement, and Hardware-Software Co-Design. Key breakthroughs include the "Transformerization" of SNNs via addition-only attention $(A^{2}OS^{2}A)$, the emergence of Saliency-Based Spiking LLMs (SpikeLLM), and the theoretical re-grounding of training dynamics on Riemannian Manifolds (MSG). By detailing these advancements, this paper delineates the current performance-oriented neuromorphic frontier and identifies the critical need for unified, robust, and hardware-aware optimization frameworks.