[edit]
DCVLR: Data Curation for Vision Language Reasoning Competition
Proceedings of the NeurIPS 2023-2025 Competition Track, PMLR 276:53-68, 2026.
Abstract
At NeurIPS 2025, we presented the first data-centric, open-data and open-models competition for data curation in vision-language reasoning. The competition focused on improving visual reasoning capabilities of vision-language models (VLMs) through data curation under a fixed supervised-finetuning (SFT) configuration and budget. Submissions were evaluated by fine-tuning a fixed VLM on the curated datasets and measuring performance on reasoning benchmarks of multiple-choice visual reasoning questions spanning diverse sources. The competition provided all necessary resources, including fine-tuning scripts, evaluation code, and baselines. Additionally, as a first of competitions of this kind, organizers also ran the entire SFT and evaluation pipeline on behalf of the contestants, thereby removing a key barrier to entry present in previous data-centric competitions. The result was the largest data-centric competition to date, with 130 signups and 20 submissions. DCVLR centered data curation as the key variable for scientific progress in VLM adaptation, highlighting its importance for real-world applications such as education, biomedicine, and scientific reasoning.