Learning to plan. Optimizing at test time.
Visiting Researcher at Carnegie Mellon University
Final-year PhD Candidate at National Yang Ming Chiao Tung University
I am currently a Visiting Researcher at WAVLab, within the Language Technologies Institute (LTI) at Carnegie Mellon University (CMU), hosted by Prof. Shinji Watanabe.
I am also a final-year PhD candidate in the Institute of Computer Science and Engineering at NYCU, advised by Prof. Wen-Chih Peng and Prof. Ping-Chun Hsieh. I am affiliated with the Advanced Database System Laboratory (ADSL) and the RL & Bandits Lab (REAL).
My research focuses on (1) test-time optimization for LLMs, (2) reinforcement learning, and (3) agentic AI applications. I am especially interested in using feedback and sequential decision-making to improve model behavior at inference time. I also maintain a strong interest in the cross-disciplinary applications of sport data science to badminton. Outside research, I find joy in playing tennis.
I am actively seeking Research Scientist and Applied Scientist positions in test-time optimization, reinforcement learning, and agentic AI. I expect to complete my PhD in September 2027. Contact me about opportunities.
WebGen-V Bench, developed during my Sony internship, was published at KDD 2026.
I joined CMU WAVLab as a Visiting Researcher, hosted by Prof. Shinji Watanabe.
A-MPQC, developed during my Sony internship, was accepted at ICML 2026.
I completed my research internship at Sony Group Corporation in Tokyo, working on multimodal agents and visual design.
TMPC, our work on test-time LLM alignment in collaboration with NVIDIA, was accepted at ICLR 2026.
SEGALE, our work on book-length translation evaluation in collaboration with NVIDIA, was accepted to the EMNLP 2025.
I joined Sony Group Corporation as an LLM Research Intern.
TMPC improves LLM outputs at inference time through short-horizon planning and reusable subgoals, with applications to translation, long-form generation, and program synthesis.
A-MPQC plans clarification questions over multiple turns to improve visual design alignment while reducing the estimated cognitive burden on users, without retraining.
SEGALE extends automatic translation metrics to book-length documents through sentence segmentation and alignment, enabling evaluation beyond fixed sentence boundaries and metric context limits.
A realistic badminton simulator with lifelike opponents and physics-based gameplay for strategy validation and tactical analysis. My ongoing badminton research also explores LLM-based agents and language-based strategy simulation.
* Equal contribution. badmintonGPT is linked as a related badminton project.
Carnegie Mellon University · WAVLab · LTI
Hosted by Prof. Shinji Watanabe. Research on test-time optimization for speech foundation models, multilingual automatic speech recognition, and speech translation.
Visit expected to conclude in May 2027.
Sony Group Corporation
Developed multimodal and multi-agent systems for digital design, including structured webpage generation and evaluation, model predictive questioning, and visual prompt discovery.
National Yang Ming Chiao Tung University · Advanced Database System Laboratory (ADSL) & RL & Bandits Lab (REAL)
Research on reinforcement learning, imitation learning, LLM alignment, long-form translation evaluation, and badminton strategy simulation. Collaborate with undergraduate and master's students on applied machine learning research.
Taiwan Semiconductor Manufacturing Company · AI4BI
Worked on LLM-based document intelligence, tabular extraction from financial documents, and agentic and decomposed information retrieval for financial report generation.
Advisors: Prof. Wen-Chih Peng and Prof. Ping-Chun Hsieh
Advisor: Prof. Wen-Chih Peng
I am seeking Research Scientist and Applied Scientist positions focused on test-time optimization, reinforcement learning, and agentic AI. I also welcome research collaborations in these areas and sport data science.