About Me

Hey there👋🏻! I’m Zian(Andy) Zheng, an AI enthusiast on a quest for challenges and unexplored horizons.

• As a CS MMath student at University of Waterloo, I’m always eager to push boundaries and embrace novel ideas.

• Beyond the study, I thrive on adrenaline-fueled outdoor pursuits like skydiving, scuba diving, free diving, climbing, and kayaking.

Whether it’s conquering algorithms or braving the wild, I embrace every chance to grow, learn, and adapt.

Research Interests

I’m generally interested in LLMs, VLMs, and AIGC. Here are a few topics I’m exploring — and what I’ve done so far:

1. Making AI models(LLMs / VLMs / Diffusion) more efficient

• OpenMoE: Explore training dynamics of routing mechanism of Mixture-of-Experts (MoE) models, aiming for better load balancing during inference.
  → Findings: context-independent specialization, early routing convergence, and end-stage drop.

• AdaVocab: Sparse vocabulary activation for next-token prediction speeds up SLM inference.

• TheMatrix: (my contribution) Inference optimization for Real-time, interactive video generation with diffusion models. [demo] → This could be the future game engine(without rewards), or an efficient video world model with them :)

2. “Simulate with AI, learn from AI, and deploy in the real world”

A vision I’m passionate about: Simulate any environment with world models, supervise policy learning with AI feedbacks(e.g. VLMs), and finally generalize to the real world.

Roadmap (in My Mind):

• ✅ Built action-controllable world models (e.g. TheMatrix, Cosmos)
• ✅ Wrapping world model + VLM reward into a Gym-style RL env (Colab Demo)

Challenges I’m working on:

• Fast world model inference: Pipelined and parallelized DiT, VAE, and post-processors in TheMatrix project.
• AI-generated rewards: Can VLMs give direct feedback, or label preferences for training reward models? Still exploring.

3. Making Human–AI interaction better

Not just research — I like to build tools too! As a heavy LLM/VLM user, I hope to build better interfaces to collaborate with AI.

Some thoughts:

• Let users segment, reuse, and compose dialogue context more effectively
• Turn chat history into a personalized knowledge base — for strengthening or sharing knowledge

Miscellaneous

• Fun academic trivia: I’m Newton’s 18th-gen academic great-great-…-grandstudent. The academic family tree is here. Sadly, I don’t think I inherited much any of his math genius — sorry, academic ancestors 😅
• I am an extreme sports lover and adrenaline junkie. You can call me ‘Tri-diver Andy’ (skydiver, freediver, scuba diver).
• I like writing poems (in English/Chinese). For example, here is one of my poems about life and love.