I am a PhD student in Computer Science and Engineering at The Chinese University of Hong Kong, advised by Prof. Yu Cheng. My research focuses on multimodal foundation models, particularly unified vision-language models, world/action models, and real-time neural rendering.

Before joining CUHK, I received my M.S. in Computer Science from the National University of Singapore and my B.S. in Computer Science from Shanghai Jiao Tong University, where I was advised by Prof. Junchi Yan.

Selected Research

Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game Worlds

I led the development of Matrix-Game 4.0, which combines executable game engines with autoregressive neural rendering to turn low-poly game visuals into detailed, interactive worlds. The system enables 480p, 60 FPS gameplay on a single RTX 5090, with 112 ms key-press-to-display latency.

Matrix-Game 3.5: Enhancing Real-Time Streaming Interactive World Models with Patch Memory

A long-horizon interactive world model with persistent memory. I led the design and implementation of its memory module, including dynamic object filtering and object-token conditioning for scene and protagonist consistency.

Cross-Task Generalization Between Understanding and Generation in Unified Vision-Language Models

BMVC 2026. We systematically study how visual understanding and generation benefit each other in unified VLMs, and identify input-output visual-space alignment as a key factor in cross-task knowledge transfer.

Less Is More: Vision Representation Compression for Efficient Video Generation with Large Language Models

AAAI 2026. We compress visual representation sequences by 4x, achieving a 10x inference speedup while reducing memory consumption and improving autoregressive video generation quality.

CLIP-MoE: Towards Building Mixture of Experts for CLIP with Diversified Multiplet Upcycling

EMNLP 2025. We upcycle complementary CLIP models into a mixture-of-experts vision encoder with minimal computational overhead, improving performance across downstream MLLM benchmarks.

News

• October 2026: We released Matrix-Game 4.0, enabling real-time neural rendering for playable games on consumer GPUs. Code, models, and the technical report are publicly available.

• 2026: Our work on cross-task generalization between understanding and generation in unified vision-language models was accepted to BMVC 2026.

• 2026: Matrix Game 3.5 was released.