Gaoxiang Luo Advances Efficient Generative AI Research
Department of Computer Science & Engineering PhD candidate Gaoxiang Luo is advancing generative artificial intelligence (AI) with research focused on improving its efficiency. His work has earned recognition at two top conferences: the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP) and the 2026 Conference on Computer Vision and Pattern Recognition (CVPR).
“My goal is to develop new generative modeling methods that enable reliable generation while also minimizing the cost to improve data efficiency,” Luo said.
In 2025, Luo presented his paper “COM-BOM: Bayesian Exemplar Search for Efficiently Exploring the Accuracy-Calibration Pareto Frontier,” written with Assistant Professor Aryan Deshwal, at EMNLP.
The paper, Luo explained, addressed an issue in selecting examples for prompting Large Language Models (LLM): researchers often optimize solely for performance. Luo’s work reframes the task as a multi‑objective optimization problem, aiming to maximize both accuracy and reliability. He developed a Bayesian optimization approach that explores the accuracy-reliability pareto-frontier in a sample-efficient manner while reasoning under noisy observations.
Luo’s second paper, "Flow Matching for Multimodal Distributions," which he worked on with Assistant Professor Ju Sun, has been accepted to CVPR 2026.
In this work, he developed an algorithm that improves three aspects of efficiency in image generation by using the intrinsic structure already present in the data. Compared to traditional training approaches, his algorithm can train models 30 times faster, while also improving inference efficiency and data efficiency.
“The model people train using our method can generate new images five times faster,” Luo explained. “And in a data-limited setting, classic training recipes will not give you good performance — but our method will still enable you to train a performing generative model.”
Improving efficiency, Luo noted, has broad implications.
By using the built-in patterns in data, less data or computing power is needed to train a generative model, Luo explained. Thereby, lowering the overall environmental impact.
Improving data efficiency could also make generative AI models more accessible, Luo said. By requiring less data, people may be able to train their own models on personal devices using only the datasets they already have.
For Luo, one of the most rewarding parts of presenting is the opportunity to engage with others in the field.
“I’m very thrilled and happy to see my work being read carefully by my peers,” he said. “Reviewers take the effort reading our paper and providing constructive feedback. I really enjoy and benefit from that process.”
Beyond his conference work, Luo recently led a team of four University of Minnesota students to an overall first place at the AgentDS-Benchmark Competition on human-AI collaboration. The competition challenged participants to solve real-world data science tasks, rewarding not only accurate solutions but also strategic and efficient use of AI tools.
Luo began doing research while he was an undergraduate at the University of Minnesota. After taking his first machine learning course with Sun, he asked to audit the professor’s lab group meetings.
“He was really supportive,” Luo recalled. “After a semester, he offered me an undergraduate research assistant position. He said he wanted me to enjoy doing research, to enjoy the process.”
For his first project, Luo developed a chest X-ray prediction system to estimate the likelihood that a patient had COVID-19. The system was later implemented at M Health Fairview. The experience showed him how AI research can make a tangible difference – something that has motivated him ever since.
Learn more about Luo's work on his website.