Seven University of Hawaiʻi at Mānoa students just pulled off a serious upset, beating teams from major tech firms and universities to win an international data curation contest. Working from a compact 1,000 example dataset, they managed to push a popular open source vision language model’s accuracy up by more than seven percentage points. The performance earned them first place, a $3,000 cash prize and a coveted podium presentation slot at the NeurIPS conference this December.
How the challenge worked and the upset
The Data Curation for Vision Language Reasoning (DCVLR) competition tasked entrants with building instruction tuning datasets, up to 10,000 examples, to improve a fixed evaluation model. UH Mānoa’s submission in the small track relied on just 1,000 examples yet produced the largest accuracy gain, about a 7.28 percentage point jump, bringing the evaluation model to roughly 46.12% accuracy on the official leaderboard, according to DCVLR.
The students and their course
The seven member squad emerged from ECE 605, a Mānoa course on large scale AI, and included Yosub Shin, Michael Buriek, Boris Sobolev, Pavel Bushuyeu, Vikas Kumar, Haoyang Xu and Samuel Watson, per University of Hawaiʻi System News. Team leader Yosub Shin said “good ideas matter just as much as computing power,” while ECE Assistant Professor Igor Molybog praised the students for showing “curiosity, discipline, and a willingness to experiment” as they developed the dataset.
What it means for Hawaiʻi tech
The result fits into a broader push to build AI capacity in the islands, with campuses expanding coursework, computing resources and local partnerships to keep talent in Hawaiʻi. One recent example: UH Hilo received a $1.4 million award this year to support open, multimodal AI research and training, signaling growing state investment in locally based AI expertise, according to UH Hilo.
Next up: NeurIPS presentation and paper
As top finishers, the Mānoa students are scheduled to present their curation methods at NeurIPS in San Diego this December and will be listed as co authors on the competition’s official publication, according to DCVLR. The conference schedule lists competition sessions in early December, and organizers say they will reproduce training and evaluation on standardized infrastructure and make the winning datasets publicly available, underscoring the contest’s open science aims, as reflected on NeurIPS.
Beyond the prize money, faculty say the exposure and co authorship give students practical credibility when applying for internships, graduate programs and research collaborations. The organizers designed DCVLR to reward careful data curation over brute force compute, a lesson Hawaiʻi instructors hope will help future cohorts compete on the world stage, according to University of Hawaiʻi System News.
