Research map / Architectures and efficiency / Spiking networks
Distribution-Aware Multi-Granularity Phase Coding: Towards Lower Conversion Error for Spike-Driven Large Language Models
Hanyuan Zheng, Haozhen Zhang, Tianshuo Chen, Zhaogeng Liu, Yi Chang, Bin Gu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add neuron training with multi-granularity phase coding over LAS (untrained special case) | LAS · LLaMA-2-7B T=8 | avg PPL: 31.73 -> 7.16 | Improved |
| add neuron training over LAS | LAS · LLaMA-3-8B T=8 | avg PPL: 8.16 -> 7.22 | Improved |
| use training-free conversion instead of SpikeLLM's training | Ours vs SpikeLLM · LLaMA-2-7B | time cost: 5h 54m -> 2m 01s | Improved |
| increase timestep T from 8 to 10 | Ours (Grain=2) · LLaMA-2-7B | avg PPL: 7.16 -> 5.73 | Improved |
| use multiple granularities (3) instead of one (ablation) | Ours · LLaMA-2-7B T=8 | avg ACC: 67.03 -> 67.31 | Improved |
| use multiple granularities (2) instead of one (ablation) | Ours · LLaMA-2-7B T=8 | avg PPL: 7.01 -> 7.16 | Hurt |
| set number of granularities equal to T (ablation) | Ours · LLaMA-3-8B T=8 | avg ACC: 71.40 vs 71.48 | No clear effect |
| add alternating optimization instead of joint (ablation) | Ours · LLaMA-3-8B 6/3 | ArcC: 47.18 -> 49.23 | Improved |
| add alternating optimization instead of joint (ablation) | Ours · LLaMA-2-7B 8/2 | Wikitext2 PPL: 6.58 -> 6.71 | Hurt |
| decouple h and theta (ablation) | Ours · LLaMA-2-7B 8/2 | Wikitext2 PPL: 7.62 -> 6.71 | Improved |
| decouple h and theta (ablation) | Ours · LLaMA-3-8B 6/3 | Wikitext2 PPL: 7.63 -> 8.53 | Hurt |
| quantize weights to 4 bit | Ours (Grain=2) · LLaMA-2-7B T=8 | Wikitext2 PPL: 7.56 -> 8.91 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SpikedAttention — SpikedAttention: Training-Free and Fully Spike-Driven Transformer-to-SNN Conversion with Winner-Oriented Spike Shift for Softmax Operation (NeurIPS 2024) · better on 6 of 6 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SpikeLLM: Scaling up Spiking Neural Network to Large Language Models via Saliency-based Spiking (ICLR 2025 · similarity 0.86)
- Temporal Misalignment in ANN-SNN Conversion and its Mitigation via Probabilistic Spiking Neurons (ICML 2025 · similarity 0.86)
- Faster and Stronger: When ANN-SNN Conversion Meets Parallel Spiking Calculation (ICML 2025 · similarity 0.86)
- Learn To be Efficient: Build Structured Sparsity in Large Language Models (NeurIPS 2024 · similarity 0.86)
- S$^2$NN: Sub-bit Spiking Neural Networks (NeurIPS 2025 · similarity 0.86)
- Inference-Scale Complexity in ANN-SNN Conversion for High-Performance and Low-Power Applications (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
