I am a Master's student in the Graduate Institute of Networking and Multimedia (GINM) at National Taiwan University, where I am fortunate to be advised by Dr. Hsuan-Tien Lin. I am currently working on machine unlearning and test-time preference alignment.
Previously, I conducted research on computer vision, multimedia, and human-computer interaction at the IGVI Lab, National Tsing Hua University, advised by Dr. Min-Chun Hu and Dr. Hung-Kuo Chu, and collaborated with Dr. Tse-Yu Pan at National Taiwan University of Science and Technology.
An evaluation of whether unlearning can return a language model to a given point in a course, scoring target, later, earlier, and other-course material separately, since target suppression or a pooled retention score can hide damage to what came before.
IGMWorld and IGMBench, an executable environment and a 110-task benchmark for coding agents that edit existing Minecraft and Terraria worlds through game modding, scored by deterministic in-game checks.
A real-time system that fuses a pianist's bodily and audio features to render affective, personalized stage visuals during live performance. An earlier version was presented as a poster at TAICHI 2025 (in Chinese).
A VR creation tool that transforms free-form creative prompts into immersive 3D environments with generative models.
A production operations platform for a two-site aesthetic-medicine clinic, used daily by 56 staff across 44 subsystems spanning attendance, leave, approvals, training, and expenses.
The approval engine compiles versioned chain templates into sequential and parallel quorum-signing steps guarded by locks and compare-and-swap, and a schema migration generalized the two-site data model to a clinics table, alongside an eight-role RBAC model and site-tagged audit events.
An AML/KYC fraud detection system for identifying suspicious activity on crypto exchanges.
Engineered 60 features and a three-layer leave-one-out toxicity scoring mechanism with graph-based features, then trained a LightGBM model with focal loss to handle severe class imbalance (3.5% positive rate).
A multimodal Dueling DRQN agent for Minecraft zombie combat that substantially outperforms visual-only DQN baselines in kill rate and survival time.
The agent fuses CNN visual features with structured spatial vectors through an LSTM and runs in Project Malmo.
A fully decentralized evolution strategies framework that achieves 100% success on Rastrigin-32 across ring, torus, and complete network topologies.
Each node maintains its own parameter vector and communicates only with graph neighbors via doubly stochastic Metropolis–Hastings mixing. Centralized and independent ES fail on the same benchmark, while a staleness-weighted asynchronous mixing rule recovers synchronous solution quality at asynchronous speed under straggler heterogeneity.
Workshops, forums, and summits I attended.