Undergraduate Student · Tsinghua University
Building language systems that reason, plan, and act with people.
I'm Zixuan (Alex) Wang, a Physics undergraduate at Tsinghua University working on AI Agents and Human-centered AI with special interests in alignment and agentic reinforcement learning, currently a research assistant at CMU supervised by Andrea Zanette.
I am Zixuan (Alex) Wang, a Physics undergraduate at Tsinghua University (class of 2023). My research aims toward human-centered AI and autonomous agentic systems — LLMs that reason, plan, and act with people.
I am a remote research assistant at Carnegie Mellon University (ECE), advised by Prof. Andrea Zanette. During my Fall 2025 exchange at UC San Diego, I was an undergraduate researcher at the UCSD MixLab (HDSI) with Dr. Zhen Wang. Earlier, I was an AI intern at MiroMind AI, mentored by Dr. Yuntao Chen.
I work on LLM algorithms that enable more reliable, adaptive, and autonomous interaction with humans — leveraging synthetic data, structured supervision, and preference-based training to improve models' understanding of human intent, while ensuring privacy and safety.
Agentic LLM Systems
Long-context reasoning, planning, tool use, and self-refinement in real-world grounded environments.
LLM Alignment
Synthetic data, structured supervision, and preference-based training for human intent — with privacy and safety built in.
RL for LLMs
Training agents for sustained multi-turn interaction and long-horizon task completion.
Papers
Submission
Mind2Dialogue: Shared-Mind Simulation and Privileged Distillation for Personalization and Theory of Mind
Submission
Lightweight Neural Refinement for Drift Calibration in Eye Tracking Systems
Submission
MEMPROBE: Probing Long-Term Agent Memory via Hidden User-State Recovery
Technical Reports
Technical Report
MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents
Technical Blog
Efficient Large-scale RL for Deep Research Agent: Rollout Chunking with Context Compression
Course Projects
Course
An Optimization View of DP LLM Fine-tuning: When Does Bias Correction Help, and Can the Optimizer Be Improved?
Course
Ego-embodied Reasoner: Egocentric Embodied Reasoning and Planning with MLLM via Reinforcement Learning
Course
EgoHOI: Prior-Guided 3D Hand-Object Interaction Reconstruction from Monocular Egocentric RGB Video
Course
Unconditional and Image-conditioned 3D Generation
Course
Human Skeleton and Skin Generation
Course
3D Visual Computing Course Notes
Notes on 3D visual computing, including geometry processing, rendering, and 3D reconstruction techniques.
Machine Learning Course Notes - Learning Theory
Comprehensive notes on learning theory, covering PAC learning, VC dimension, and statistical learning foundations.
I build with agents, and agents run on tokens. Every chart below is live telemetry from my own machines — every token my coding agents have read, cached, and written.