Task-Guided ANN-to-SNN Conversion for Large Language Models via Sparse-Burst Phase Encoding

Converting pretrained large language models (LLMs) from artificial neural networks (ANNs) to spiking neural networks (SNNs) provides an efficient way to reduce inference energy without costly training from scratch. However, accurate ANN-to-SNN conversion with few timesteps remains challenging, as activation outliers in a small number of ANN channels expand the dynamic range and require finer resolution in SNNs for accurate spike-based representation. Existing studies have made progress, but still face two limitations: some improve representation fidelity by adding timesteps or neurons, increasing inference overhead, while others optimize task-agnostic reconstruction, limiting task performance. To tackle this, we propose a novel conversion method that combines an efficient spike encoding scheme with a task-guided conversion framework. First, we introduce Sparse-Burst Phase (SBP) encoding to improve representational efficiency. SBP uses an integer resolution parameter together with signed burst spikes to decouple activation resolution from timestep count, enabling flexible representation of wide and nonuniform LLM activations without additional timesteps or neurons. Building on SBP, we further develop Task-Guided Adaptive-Resolution Conversion (TARC) framework to improve task performance. TARC first allocates the representation budget across channels according to their task sensitivity and then aligns the converted SNN with the pretrained ANN through channel-wise threshold and offset calibration. Extensive evaluations are conducted on seven LLaMA and Qwen models ranging from 7B-70B, covering language modeling, commonsense reasoning, and knowledge-intensive tasks. With only 4 timesteps, TARC achieves state-of-the-art performance among converted SNNs while closely matching the original ANN models. These results demonstrate that TARC enables accurate and efficient spike-driven LLMs under limited temporal and hardware budgets.