PhD candidate · 2027 / AI researcher
I am a PhD candidate in Computer Science and Artificial Intelligence at Oregon State University, advised by Liang Huang and co-advised by Stefan Lee. At Genies, I worked on GRPO post-training, text-to-SQL evaluation, and diagnosing failures in multi-turn tool use; at Amazon, I researched multimodal generation and cross-lingual retrieval. My PhD research connects synthetic data, multilingual generation, and streaming evaluation through sign-language translation and 3D human motion.
Seeking full-time Research Scientist, Applied Scientist, Research Engineer, and ML Engineer roles starting in 2027.
01 / Selected work
Built direct ASL/CSL/DGS translation without a spoken-text pivot at inference and created 112K cross-lingual pairs through back-translation. On held-out back-translated inputs across six directions, reduced mean motion error (DTW-PA-MPJPE) by 19% and achieved 2.3× faster single-GPU inference than a three-stage cascade.
Synthetic data & curation · multilingual generation · mBART · VQ-VAE
Built streaming sign-to-sign generation and proposed CA-Stream-AL to measure computation-aware latency. Reduced latency by 38% versus full-sentence decoding, with 9% higher motion error (DTW-PA-MPJPE) and 2.1 points lower BLEU-4.
Evaluation design · streaming inference · wait-k · multi-path training
Built variable-length 3D sign generation with geometry-aware objectives for articulated hands and consistent skeletons. Reached 14.22 back-translation BLEU-4 on PHOENIX14T development data, 3.08 points above the XLM-R baseline.
3D human motion · XLM-R · OpenPose · inverse kinematics
02 / Industry
2024–2025
Built an LLM user-insight application and ran GRPO post-training experiments for text-to-SQL. Turned product requirements into evaluation cases that check semantic correctness beyond successful SQL execution. Diagnosed model, tool, and execution-environment failures from multi-turn database/SQL traces, refined interfaces and feedback, and reran failing and previously passing cases. Built reproducible data, distributed-training, checkpoint, and evaluation workflows with DeepSpeed, Hugging Face Accelerate, and Docker.
2023
Researched generative models for long multimodal sequences and cross-modal dependencies.
2022
Developed multilingual representation-learning methods on XLM-R for cross-lingual retrieval.
2019–2020
Built generative sequence models for music style transfer and MIDI generation.
03 / Writing
Simultaneous Translation between Sign Languages
Zetian Wu, Bowen Xie, Stefan Lee, and Liang Huang · arXiv:2609.35608
Direct Translation between Sign Languages
Zetian Wu, Bowen Xie, Wuyang Meng, Milan Gautam, Stefan Lee, and Liang Huang
Geometry-Aware Losses for Structure-Preserving Text-to-Sign Language Generation
Zetian Wu, Tianshuo Zhou, Stefan Lee, and Liang Huang
Inducing Generalizable and Interpretable Lexica
Findings of EMNLP
MultiBench: Multiscale Benchmarks for Multimodal Representation Learning
NeurIPS Datasets and Benchmarks
Interactive Rare-Category-of-Interest Mining from Large Datasets
AAAI
Education. PhD candidate, Oregon State University (expected 2027) · MSE, Johns Hopkins University · BS, Zhejiang University.
Tools. Python, SQL, PyTorch, Hugging Face Transformers & Accelerate, DeepSpeed, and Docker.