CV

Education

University of Electronic Science and Technology of China (UESTC)

Computer Science & Technology, Direct-entry Ph.D., 2026.03 - 2029.09
Computer Technology, M.S., 2024.09 - 2026.03
Artificial Intelligence, B.S., 2020.09 - 2024.07
Supervisor: Prof. Malu Zhang
National Scholarship; National College of AI (Shenzhen Hetao College), Outstanding Camper

SpikingLM: Towards Fully Spiking Language Model, ICML 2026 (2025.05 - 2025.10)

Motivation: Extending SNNs to modern language models faces two challenges: dead neurons in deep networks cause gradient degradation, and removing Softmax for efficiency weakens token-wise competition.
Method: Proposed the first fully spiking language model. Designed Distribution-aware Scaling (DAS) to suppress dead neurons with zero inference overhead, and Spike2Max attention that restores winner-takes-all via base-2 exponentiation and max-subtraction using only shift operations.
Results: Spike2Max cuts energy by over 95% versus Softmax; SpikingLM reduces overall energy by 57.9% and achieves SOTA on the GLUE benchmark.

BSO: Binary Spiking Online Optimization Algorithm, ICML 2025 (2024.12 - 2025.05)

Motivation: Training binary SNNs relies on full-precision latent weights to accumulate gradients, incurring huge memory overhead that limits deployment on edge devices.
Method: Proposed the Binary Spiking Online (BSO) algorithm and its temporal-aware variant (T-BSO), which directly update binary weights via flip signals, fully eliminating the storage dependence on latent weights.
Results: Outperforms SOTA by 4.5% on ImageNet while greatly reducing training memory, enabling efficient training on neuromorphic hardware.

Towards Accurate Binary SNNs: Learning with AGMM, AAAI 2025 (Oral) (2024.07 - 2025.01)

Motivation: Binary SNNs suffer from frequent weight sign flipping during training, which destabilizes the update direction, hinders convergence, and degrades performance.
Method: Proposed the Adaptive Gradient Modulation Mechanism (AGMM); theoretically linked flipping frequency to gradient mean and variance, and designed a temporal-aware modulation function that adaptively scales gradients.
Results: Achieves SOTA across static-image and neuromorphic datasets, narrowing the accuracy gap between binary and full-precision SNNs.

ModelBest (Mianbi Intelligence), Multimodal Data Construction & Post-training (2025.12 - 2026.06)

Data Pipeline: Embedded 240M pretraining samples with Qwen-Embedding-2B, performed similarity-based retrieval over AI2D with deduplication, and constructed 100K high-quality scientific Hard-QA and caption samples.
Results: Post-training markedly improved scientific reasoning: MMMU Science +8.67 (Chemistry +20.00), MMStar Sci&Tech +5.60, MathVista +3.28, AI2D_TEST +0.87.

Manifold-Optimized Rotational Quantization for LLM Acceleration (2024.07 - 2025.01)

Low-bit Architecture (W4A4KV4): To tackle accuracy collapse caused by activation outliers in ultra-low-bit LLM quantization, designed a post-training quantization (PTQ) framework based on learnable rotation matrices.
Multi-strategy Co-design: Fused rotation parameters into weights for zero inference overhead, and used online Hadamard transforms for extreme compression.
Results: Validated on LLaMA-2/3 models, reducing the W4A4KV4 versus FP16 accuracy gap from over 20 points to within 2.9 points, surpassing methods such as LLM-QAT on zero-shot reasoning.

Language: CET-6 score 552; solid English academic reading and technical writing.
Deep Learning: Familiar with common deep learning algorithms; proficient in PyTorch and Tensorflow for model building and training.
Tools: Linux, Shell scripting, Git version control, and Docker containerized deployment.