👤 About
I am a final-year Ph.D. student at the ARISE Lab, Department of Computer Science and Engineering, The Chinese University of Hong Kong (CUHK), advised by Prof. Michael R. Lyu. I received my B.Eng. (First Class Honours, AI: Systems and Technologies) from CUHK in 2023.
My research focuses on making AI agents reliable on long-horizon tasks such as coding, research, and gaming, with an emphasis on multi-agent collaboration, agent memory, agent harnesses, post-training, and test-time self-evolution. Through industry roles at Tencent Games and Roblox, I have applied this research to practical problems.
I also enjoy building imaginative worlds and interactive experiences such as games; a few of them live in my Collection.
🗂️ Collection yxwan123.github.io/collection ↗
💼 Experience
- Experimenting with model architectures and training methods to improve prediction performance over the production baseline
- Owning the offline experiment loop — data preparation, training, evaluation, error analysis — in Roblox's production ML stack
- Two-level LLM planner with an episodic memory module for AI teammates; rescue skill refactored into a code-controlled state machine
- Structured-prompt chat module with dropout filtering; multi-voice TTS fine-tuned on four voice samples
- Game-config agent rebuilt as a code-led pipeline: prompt 65K → 909 chars, ~3× lower latency, error rate ~20% → ~3%
- Designed and developed reflection and voting modules for the Cognitive-Pro Deep Research Agent
- Achieved +6% performance improvement on research agent benchmarks
- Co-authored technical report (arXiv) and published paper (arXiv 2601.15808)
