Kaiwei Che

Kaiwei Che

Ph.D. student at Peking University. Research in Spiking Neural Networks, brain-inspired algorithms, and LLM quantization.

About Me

I am Kaiwei Che (Richard), a third-year Ph.D. student at Peking University (2023–2027), advised by Prof. Yonghong Tian and Prof. Li Yuan. My research focuses on AI Infra and Efficient AI, including model sparse quantization, training and inference acceleration, and brain-inspired algorithms. I have published 3 CCF-A first-author papers at ICML (Spotlight), NeurIPS (Spotlight), and AAAI, along with 2 SCI journal papers, with 360+ total citations.

Before this, I received my Master’s degree from Southern University of Science and Technology (SUSTech) in 2023, advised by Prof. Qinghu Meng. I received my Bachelor’s degree from Shenzhen University in 2020.

I have interned at Huawei (2012 Lab, Research Intern, 2021–2023) and DJI (RoboMasters, 2018), with extensive experience in algorithm research and engineering practice.

Research Interests

  • AI Infra & Efficient AI: model sparse quantization, training/inference acceleration, brain-inspired algorithms
  • Spiking Neural Networks (SNNs): energy-efficient deep learning, neuromorphic computing
  • LLM/VLM: post-training alignment (SFT, RLHF), RAG, agent systems
  • Event-based Vision: neuromorphic sensing, efficient event processing

Publications

First Author

Co-Author

  • Neural Networks 2025 Spatially-enhanced Spiking Neural Network for Efficient Point Cloud Analysis
    Yijie Lu, Zhiyi Pan, Renrui Zhang, Yanhao Jia, Kaiwei Che, Zhaokun Zhou

  • NeurIPS 2024 Spiking Transformer with Experts Mixture
    Zhaokun Zhou, Yijie Lu, Yanhao Jia, Kaiwei Che, Jun Niu, Liwei Huang, Xinyu Shi, Yuesheng Zhu, Guoqi Li, Zhaofei Yu

  • TIP, under review Spikformer V2: Join the High Accuracy Club on ImageNet with an SNN Ticket
    Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, Li Yuan

  • TNNLS 2024 Accurate and Efficient Event-based Semantic Segmentation Using Adaptive Spiking Encoder-Decoder Network
    Rui Zhang, Luziwei Leng, Kaiwei Che, Hu Zhang, Jie Cheng, Qinghai Guo, Jianxing Liao, Ran Cheng

  • IEEE TCDS 2024 Automotive Object Detection via Learning Sparse Events by Spiking Neurons
    Hu Zhang, Yanchen Li, Luziwei Leng, Kaiwei Che, Qian Liu, Qinghai Guo, Jianxing Liao, Ran Cheng

  • IEEE TCDS 2024 Spatial-Temporal Spiking Feature Pruning in Spiking Transformer
    Zhaokun Zhou, Kaiwei Che, Jun Niu, Man Yao, Guoqi Li, Li Yuan, Guibo Luo, Yuesheng Zhu

  • ICANN 2024 A Multi-modal Spiking Meta-learner with Brain-Inspired Task-Aware Modulation Scheme
    Jun Niu, Zhaokun Zhou, Kaiwei Che, Li Yuan

  • CVPR 2022 Discrete Time Convolution for Fast Event-Based Stereo
    Kaixuan Zhang, Kaiwei Che, Jianguo Zhang, Jie Cheng, Ziyang Zhang, Qinghai Guo, Luziwei Leng

  • ROBIO 2021 Motion Planning for Hexapod Robots in Dynamic Rough Terrain Environments
    Bingyi Xia, Kaiwei Che, Zhilong Tang, Jiankun Wang, Max Q-H Meng

Research Projects & Experience

AI Infra & Efficient AI

  • NPUSlim: LLM Quantization, Sparse & Acceleration Framework on Huawei NPU (Provincial Key Project)
    A framework for quantization, sparsity, and inference acceleration on Huawei NPUs. Supports mainstream model integration and deployment via vLLM / vLLM-ascend. Implements full-pipeline 2:4 structured sparse operator adaptation on NPU, integrates GPTQ, SparseGPT and other PTQ algorithms. Features a layer-block streaming quantization pipeline with trillion-parameter-scale offline quantization capability on a single machine.

  • Winner-Take-All Spiking Transformer for Language Modeling
    Proposes WTA-based attention to replace softmax for fully spike-driven, floating-point-free computation. Develops both Decoder-only and Encoder-only SNN Transformers, achieving SNN SOTA across multiple datasets with only 1/14 the energy of Qwen-1.5B.

Agent

  • SF-RAG: Selective Filtering for Retrieval-Augmented Generation
    Addresses context overload and intermediate information issues in traditional RAG. Proposes a lightweight Reasoner for parallel inference and CoT chunk-level utility scoring, enabling dynamic retrieval document filtering. Outperforms SELF-RAG, SURE, and other baselines on 5 open-domain QA benchmarks. The Reasoner is an LLaMA-8B model fine-tuned via SFT on GPT-4o synthetic data.

  • Personal Time Management Agent
    A multi-workflow agent built on LangGraph that routes weekly/daily planning and review workflows by intent. Uses interrupt/resume for user confirmation and information completion, constraining the LLM from unauthorized plan modifications for improved controllability.

Technical Blogs

  • LLM/VLM Technical Blog
    Systematically covers VLM technical directions, including discriminative models (CLIP, BLIP-2, Flamingo, Qwen-VL) and generative models (DDPM, LDM, DALL-E, DiT). Includes multimodal fine-tuning experiments using LLaMA-Factory with Qwen3-VL SFT.

  • RLHF Technical Blog
    Covers RL alignment methods (PG, TRPO, PPO, GRPO) from both mathematical principles and engineering implementation perspectives. Includes multimodal RL alignment experiments on Geo3K using verl with Qwen2.5-VL and LoRA GRPO.

Academic Service

I have experience in reviewing for high-impact AI conferences and journals, including:

Conferences: NeurIPS, ICML, ICLR, CVPR, AAAI

Journals: TIP, TNNLS, Pattern Recognition, Scientific Reports

Skills

  • Programming Languages & Frameworks: Python, C++, Triton; PyTorch
  • Post-training & Alignment: SFT, RLHF (PPO, GRPO); Frameworks: LLaMA-Factory, verl
  • Agent Systems: RAG, Tool use; Framework: LangGraph
  • Inference Optimization: Quantization, sparsity, inference acceleration, NPU deployment; Framework: vLLM
  • Open Source: Spikingjelly (2k+ stars), NPUSlim, RoboRTS (800+ stars)
  • Languages: CET6

Contact

  • E-mail: chekaiwei@stu.pku.edu.cn
  • Google Scholar: Kaiwei Che