Achieve Latency-Efficient Tempora-Coding Spiking LLMs via Discretization-Aware Conversion

ICLR 2026 Conference Submission16048 Authors

Keywords: Large Language Models, Spike Neural Network, Temporal Coding, Ann-to-snn Conversion

Abstract:

Large language models (LLMs) have achieved remarkable success while introducing critical energy bottlenecks that challenge sustainable deployment. Spiking neural networks (SNNs) provide a promising approach for energy-efficient spiking LLMs via ANN-to-SNN (A2S) conversion. Among various spike coding methods, time-to-first-spike (TTFS) coding is particularly appealing as it conveys information with a single spike, further reducing energy consumption. However, existing TTFS-based A2S conversion relies on continuous-time assumptions, requiring prohibitively large latencies (e.g., 4096 time steps) to approximate ANN's continuous values.
This dependency leads to unacceptable inference delay in deep models, particularly LLMs, posing significant challenges for developing practical temporal-coding spiking LLMs. In this paper, we propose a discretization-aware theoretical framework that establishes a precise correspondence between discrete TTFS-based SNNs and ANNs. Our key insight reveals that conversion errors are bounded by latency-dependent terms. Motivated by these, we introduce the Quantization-Consistent ANN-to-SNN (QC-A2S) conversion, which integrates low-bit quantization with discretization-compatible TTFS neurons, achieving latency-efficient temporal-coding spiking LLMs.
Comprehensive evaluation on LLaMA models demonstrates comparable performance with dramatically reduced latency.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 16048