A valedictorian of the elite computer science talent program at Hanoi University of Science and Technology has had papers accepted at two of the world's top conferences on machine learning and natural language processing.
Le Danh Vinh graduated with distinction in late September with a GPA of 3.92 out of 4.
He works at VinMotion, Vingroup's humanoid robotics company, and is a member of DreamAI, a research group founded by Trung Le, a machine learning researcher at Australia's Monash University.
Vinh said he was glad not only to graduate with top marks but also to have found a field he loves. "It's something I never imagined in my early days at the university," he said.
A former student of the specialized math class at Bac Ninh High School for the Gifted, Vinh was admitted to the university without an entrance exam in 2022 on the strength of a second prize in the national excellent student contest in mathematics. He then passed an interview to enter the talent program.
From his first year, he went beyond set textbooks and his lecturers' reading lists to study specialist material on machine learning and deep learning, subjects not formally taught until the third year.
He scored a perfect 4.0 GPA in his first two semesters and won first prize in analysis at the national Mathematical Olympiad for university students. By then he had settled on a path in artificial intelligence.
His search for internships and lab positions, which began in the second semester of his second year, went nowhere at first. His research career took shape only in his third year, when he joined ZenAI, a generative AI startup.
He led several projects there. Ngo Van Linh, a lecturer at the university who also advises the startup, gave him the chance to do research alongside him.
 |
|
Le Danh Vinh with his parents on his graduation day at Hanoi University of Science and Technology. Photo courtesy of Vinh |
Vinh's research focuses on efficient AI, particularly methods that cut the cost of running models and speed up inference, the stage at which a trained model produces its answers.
He also works on vision-language models (VLMs), which process and combine information from different types of data such as images and text.
"I'm also interested in developing more efficient models and using VLMs for vision-language tasks, as well as applications in humanoid robots," he said.
The final push on his paper for ICML 2026, the International Conference on Machine Learning, coincided with end-of-semester exams and a full workload at the company. He was also short of the computing hardware needed for experiments.
"Some days I slept four hours or less because I wanted the paper to be accepted so badly, but I couldn't neglect my work at the company or my exams at school," Vinh said.
ICML accepted the paper, on which Vinh is first author. The conference, rated A*, the top tier of the CORE ranking of computer science conferences, was held in Seoul in July and turned away more than 73% of a record of some 24,000 submissions, according to its organizers.
Vinh is also a co-author of a paper accepted to EMNLP 2026, a leading conference on natural language processing, to be held in Budapest in late October.
Linh, who has worked with Vinh for nearly two years, said he was struck by his independent thinking and persistence.
When Linh set him a new direction, knowledge distillation, a technique for transferring what a large model knows into a smaller one, applied to resource-hungry multimodal embedding models, the early results were inconclusive. Vinh kept working through the problems, Linh said.
When ICML reviewers raised tough questions, Vinh built his own arguments and verified them convincingly without waiting to be guided.
At VinMotion, he is looking for new challenges. He plans to gain more work and research experience before pursuing graduate studies.