I'm currently an undergraduate student at UBC and research assistant in the UBC NLP Group, advised by Peter West. I'm interested in using our insights towards how human think, learn and develop to improve how models generate, reason, and evolve. My recent research interests include:
- Creativity. Understanding how intelligent systems can generate genuinely novel ideas rather than converging toward consensus.
- Memory. Exploring memory as a fundamental component of intelligence, not merely as information storage.
- Non-language communication/reasoning. Investigating forms of reasoning and conversation that extend beyond human language.
Experience
I am currently pursuing a B.Sc. in Computer Science at UBC, and hold an Bachelor of Laws (LL.B.) from Fujian Agriculture and Forestry University. Prior to transitioning into computer science, I worked as a legal assistant in the China legal team at Royal DSM (now dsm-firmenich).
At UBC, I have worked across multiple research areas. My work includes co-authoring Position: Universal Aesthetic Alignment Narrows Artistic Expression (ICML 2026 Spotlight) with Khalad Hasan and Shan Du; conducting travel behavior data collection, cleaning, and analysis with Khalad Hasan and Mahmudur Fatmi; exploring computational creativity with Liane Gabora; and studying myelin and memory with Reza Khanbabaie.
Selected Projects
Position: Universal Aesthetic Alignment Narrows Artistic Expression ICML 2026 Spotlight (top 5%) · W. Guo, Q. Qian, K. Hasan, S. Du · Code & Data · Project Page · ICML Page
When users request “anti-aesthetic” outputs for artistic or critical purposes, over-aligned generative models default to conventionally beautiful results anyway — and reward models penalize anti-aesthetic images even when they perfectly match the user’s prompt. We argue this reflects a systematic bias that compromises user autonomy and aesthetic pluralism.
BCATUS: Travel Behavior Tracking Oct. 2024 – Apr. 2026
Built an iOS app (BCATUS) in Swift that passively tracked the travel behavior of 500+ users across Metro Vancouver and the Okanagan, collecting 12,600+ trips to inform government transportation policy. On the data pipeline side, we designed algorithms to impute missing trip-purpose labels (11.2% relative improvement), detect and remove erroneous indoor GPS loops, and merge fragmented transit segments — automating 95%+ of trip processing and cutting analysis time by 50%.
AI-Powered Exam Generation Platform · React · Node.js · PostgreSQL · Docker · BERT May – Aug. 2025
A microservice-based platform for automated university exam generation. I designed a semantic question recommendation service via cosine similarity over BERT embeddings, with time-versioned caching that cuts redundant computation by 85–95%. I also designed a logic-preserving shuffle algorithm using dependency chains and similarity clustering, and replaced O(n!) permutation generation with O(1) precomputed lookup for 2–5 option questions — achieving up to 24× speedup for the most common cases.