Hi, I’m Brice — an engineer and research assistant working on AI, video generation, interactive world models, and the mathematical foundations behind them.
My current research interest is centered on generative video world models: models that go beyond producing short visually plausible clips and begin to simulate persistent, controllable, and interactive visual worlds.
Currently, I am especially focused on:
- Long-horizon video generation — temporal coherence, motion modeling, and controllable visual synthesis
- Interactive world models — memory consistency, out-of-view dynamics, action control, and interaction
- Mathematical foundations for generative AI — probability, stochastic processes, optimization, and uncertainty
I am drawn to questions such as:
- How can a generative model preserve objects, scenes, and events over time?
- How should memory be represented when important state moves out of view?
- How can action and interaction turn video generation into world simulation?
Since 21 March 2026, I have been working as a Research Assistant under Xinyu Zhang at the University of Auckland, focusing on video generation and generative video world models, including directions related to LiveWorld.
I am currently pursuing a Master of Artificial Intelligence at the University of Auckland (New Zealand), with the long-term goal of continuing toward doctoral-level research in generative AI and interactive world models.
🎓 Education
B.Eng. in Electronic Information Engineering
Harbin University of Science and Technology · Aug 2021 – Jul 2025
My undergraduate training emphasized a dual foundation in systems and mathematics, including:
- Computer systems and embedded architectures
- Mathematical modeling and numerical optimization
- High-performance and parallel computing
- Compiler theory and low-level program optimization
GPA: Top 1% of cohort
Master of Artificial Intelligence (180 points)
University of Auckland · Feb 2026 – Dec 2027 (expected)
Research focus:
Video Generation × Interactive World Models × Mathematical Foundations
I am particularly interested in:
- Transformer-based, diffusion-style, and hybrid generative architectures
- Long-horizon video generation and scene-level temporal consistency
- Memory, action control, and interaction in generative video world models
- Mathematical perspectives on optimization, stability, and generalization
I intend to continue toward doctoral-level research in this direction.
🔬 Research Experience
Research Assistant — University of Auckland
Supervisor: Xinyu Zhang · Mar 21, 2026 – Present
My current research direction is centered on video generation and generative video world models. The broader goal is to understand how video models can move from short-form generation toward long-horizon simulation with memory, controllability, and interaction.
Current focus areas include:
- Long-horizon video generation — extending temporal consistency beyond short clips
- Static memory consistency — preserving scene layout, object identity, and spatial structure
- Dynamic memory consistency — modeling out-of-view events, state changes, and hidden dynamics
- Action control — conditioning generation on actions, motion, or controllable latent behavior
- Interactive ability — enabling users or agents to affect generated environments through instruction or action
Related to this direction, I follow problem-oriented research maps such as Awesome Interactive World Model, which organize interactive world model papers by the questions they address instead of only by model name or release date.
This reflects how I prefer to read the field: starting from core research problems such as static memory consistency, dynamic memory consistency, action control, interactive ability, post-training, physics, and evaluation, then asking what each line of work actually solves.
Contributions and pull requests are welcome.
💼 Industry Experience
System Technology Intern — Tencent Cloud
CSIG Division · Xingxinghai Lab · Jul 2024 – Sep 2024
Worked on production-grade cloud infrastructure with an emphasis on automation, reliability, and performance-aware system design.
Key contributions included:
- Automating cloud server deployment pipelines for stable production environments
- Improving internal system workflows to enhance efficiency and reliability
- Collecting and analyzing operational metrics to support data-driven optimization
- Authoring technical analysis reports emphasizing correctness and traceability
- Participating in discussions around distributed system reliability and
kernel-adjacent performance tuning
This experience shaped my engineering habits around automation, reproducibility, and careful debugging, which now carry into my research workflow.
🧠 Research & Technical Interests
-
Video Generation
Temporal modeling, motion-aware architectures, diffusion-based video synthesis, long-horizon generation -
Interactive World Models
Static memory consistency, dynamic memory consistency, action control, interaction, evaluation -
Deep Learning Systems from First Principles
From-scratch implementation of reverse-mode autograd, neural network layers, optimizers, data loading, and training visualization in NumPy -
Self-directed Technical Reproduction
Ability to quickly study, reproduce, debug, and document unfamiliar technical systems, including a prior SO-100 / LeRobot robotics workflow reproduction -
Probability Theory & Stochastic Processes
Measure-theoretic probability, Markov chains, SDEs, probabilistic graphical models — as rigorous foundations for generative AI -
Mathematics for AI
Optimization, numerical methods, information theory -
Systems & Rust Engineering
Async runtimes, compilers, concurrency models, performance-aware design -
Algorithms & Data Structures
Graphs, optimization algorithms, compiler IRs, complexity-aware implementations
🌲 Skill Tree
🧩 Programming Languages
C · C++ · Rust · Zig · Go · Python · Java · JavaScript · Swift
Languages are treated as tools, with emphasis placed on abstraction, correctness, and performance trade-offs.
🔬 Research & Technical Projects
(Selected research and engineering work)
| Project | Description | Keywords |
|---|---|---|
| Research Assistantship: Video World Models (Active) | Research under Xinyu Zhang at the University of Auckland on video generation and interactive world models, with attention to memory consistency, action control, and long-horizon simulation. | Video Generation · World Models · Research |
| Problem-Oriented World Model Reading | I follow and organize my reading around research problems such as static memory consistency, dynamic memory consistency, action control, interaction, physics, and evaluation, using resources like Awesome Interactive World Model as references. | World Model · Research Taste · Reading |
| MiniTorch | Built a lightweight PyTorch-inspired deep learning framework from scratch with NumPy, including reverse-mode automatic differentiation, differentiable tensor operations, neural network modules, SGD/Adam optimizers, MNIST training, computation graph visualization, and unit tests. | Autograd · Deep Learning Systems · NumPy |
| SO-100 / LeRobot Reproduction | Independently reproduced and documented an unfamiliar robotics workflow as practice in rapid self-learning, system setup, debugging, and technical documentation. | Self-learning · Reproduction · Documentation |
| Video Generation Study (Active) | Studying diffusion-based and autoregressive architectures for video synthesis, with emphasis on temporal coherence, motion modeling, and long-horizon generation. | Video Generation · Diffusion · Temporal Modeling |
| Probability Theory Study (Active) | Systematic study of measure-theoretic probability and stochastic processes as mathematical foundations for generative models and diffusion-based AI. | Probability · Stochastic Processes · Math |
🌐 Contact & Presence
- GitHub → https://github.com/BriceLucifer
- X (Twitter) → https://x.com/Bricelucifer
- Email → 2376671337@qq.com
- KnowledgeBase -> https://atlas-32q.pages.dev/
This site is built with **Hugo + PaperMod** and documents an evolving research trajectory in **video generation and interactive world models**.