Building efficient AI systems

I am Hanzhen Wang, a fourth-year undergraduate student majoring in Automation at Shanghai Jiao Tong University.

I am currently working in DAI Lab led by Prof. Guohao Dai.

My research focuses on Efficient Artificial Intelligence Models, optimizing from both algorithm and system levels for cloud and edge scenarios. My current interests include Large Language Models, Embodied AI, and Agentic AI.

LLM Embodied AI Agentic AI

Education

Shanghai Jiao Tong University emblem

September 2023 — Present · Shanghai, China

Shanghai Jiao Tong University

Undergraduate, School of Automation and Intelligent Sensing · Major in Automation

Publications

* Equal contribution · † Corresponding author

C1
ICML 2026Published

SpecPrune-VLA: Accelerating Vision-Language-Action Models via Action-Aware Self-Speculative Pruning

Hanzhen Wang*, Jiaming Xu*, Jiayi Pan, Yongkang Zhou, Guohao Dai†

C2
ASPLOS 2026Published

SpeContext: Enabling Efficient Long-context Reasoning with Speculative Context Sparsity in LLMs

Jiaming Xu, Jiayi Pan, Hanzhen Wang, Yongkang Zhou, Jiancai Ye, Yu Wang, Guohao Dai†

C3
ASP-DAC 2027To appear

MTPMoE: Exploiting Activation Skewness in Single-GPU MoE Multi-Token Prediction via Parallel Dense-Sparse Computing

Yongkang Zhou, Jiaming Xu, Jiayi Pan, Hanzhen Wang, Tianlang Zhao, Guohao Dai†

C4
2026Under review

FlashMoD: Approximate Dynamic Programming Helps Efficient Mixture of Depth in Large Language Models on Edge Devices

Jiaming Xu, Hanzhen Wang, Tianlang Zhao, Yongkang Zhou, Yifan Jiao, Chengze Yuan, Guohao Dai†

C5
2026 Preprint

DLite: History-Free Speculative Decoding for Agentic AI

Jiaming Xu*, Hanzhen Wang*, Tianlang Zhao, Jiewen Xiao, Chengze Yuan, Jiayi Pan, Yongkang Zhou, Jiancai Ye, Guohao Dai†

Internship Experience

September 2026 — Present

Research Intern @ Dots Infra

RedNote(小红书)

Supervisor: Dr. Zheng Qu

Research directions

  • Speculative Decoding Design and Training in LLM
  • Speculative Decoding Inference Framework Optimization in GPU Cluster

Project Experience

A unified framework that supports training and evaluation of DLite, DFlash, Dflash2 and DSpark.

  • Flexible dataset conversion: Supports conversion from arbitrary dataset formats.
  • Efficient offline training: Efficient data collection, reading and writing, and offline draft model training.
  • Lightweight online pipeline: A target–draft decoupled, asynchronous inference–training pipeline within a single node.
  • On-policy sampling: Eliminates numerical precision mismatches between offline data and training caused by differences in GPU models, inference backends, and kernels.
Cloud Computing

Airport Ground Support Equipment Tracking System

In collaboration with China Eastern Airlines, supervised by Prof. Zhenyu Huang.

An IoT-enabled tracking and billing system for unpowered airport ground vehicles.

  • At the Network Layer: Built a Java Spring Boot backend for high-throughput, low-latency edge-to-cloud data ingestion.
  • At the algorithm level: Developed temporal RSSI filtering to reduce multipath noise and improve positioning accuracy.
  • Cloud Database: Deployed a MySQL-based time-series pipeline to monitor vehicle locations and usage duration in real time.