Minshen Zhang

UC San Diego

Master in Computer Science (CS75)

ShanghaiTech University

Bachelor of Computer Science and Technology

Overall GPA: 3.69/4.0; Major GPA: 3.8+/4.0

English Tests: TOEFL(100), GRE(320), CET6(602), CET4(607), Duolingo(135)

UC Berkeley Extension

Berkeley GLOBE Student Program

Overall GPA: 4.0/4.0

Shanghai Alibaba Ant Group NLP Research Lab

Research Intern of NLP Research Group

Mentor: Xiang Hu

Observed that the multi-subspace learning capability of Multi-Query Attention significantly enhances the expressiveness of the attention mechanism. Given the high similarity in their mathematical formulations, we hypothesize that the MLP component in Transformer blocks can also be decomposed into multi-head computations in a similar manner.

Designed an integration method for multi-head MLPs within standard Transformer blocks and a corresponding triton algorithm that improves throughput (1.06x) and reduces memory usage (over 2x) compared with LlamaMLP (SwiGLU) baselines; project lead.

ShanghaiTech Kewei Tu's Lab

Research Assistant of NLP Research Lab

Mentor: Prof. Kewei Tu (ShanghaiTech University)

Building on Have I Learned This Entity? (ICLR 2025), using sparse autoencoders to test whether an LLM encodes entity-specific knowledge in context and to identify sparse latent directions. Applying these directions to downstream tasks such as feedback-guided RAG retrieval, RAG-based hallucination detection and latent direction steering; project lead

Shanghai Qizhi Institute

Research Intern of Machine Learning Research Group

Mentor: Prof. Tao Du (Tsinghua University)

Mathematically discovered an analytical solution for real-time inverse-refraction problems based on knowledge of physical optics, linear algebra and computer numerical methods. Developed real-time machine learning algorithms for water surface reconstruction to assist robot control tasks. Based on Snell's law, derived the relationship between water surface normal vectors and visual information, proving the feasibility of machine learning approaches. Implemented core algorithms using Nvidia-Warp-Cuda, achieving breakthrough performance (from 500ms to 8ms per frame).

Shanghai ElanTech Lab

Embodied Intelligence Research Group

Advisor: Prof. Lan Xu (ShanghaiTech University)

Implemented and improved network structures including ResNet, BiRNN, Diffusion, and Transformers. Built large-scale human motion capture facilities using OptiTrack and Z-cam camera arrays for data collection. Integrated existing algorithms with Motion Diffusion Model to reconstruct human motion using IMU sensor signals.