A Fully TTFS Spiking LLM Scales to 1.5B Parameters With at Most One Spike Per Neuron

arXiv 2609.05151·low signalXBlueskyLinkedInCopy link

arXiv 2609.05151 introduces a reference-based strategy for encoding the four LLM components that conventional time-to-first-spike coding cannot handle — embedding layers, layer normalization, attention-related operations, and dropout — and trains a fully TTFS-based spiking network end to end. On BERT and GPT-2 it matches ANN counterparts on natural language understanding and commonsense reasoning, though a clear gap remains on language modeling perplexity. The authors are explicit that the reported energy figure is a spike-count proxy under an established cost model, not a measurement on neuromorphic hardware.