DecoupleMix: Decoupled Ratio Search and Convex Allocation for Scalable VLM Data Recipes
2026-07-27 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial Intelligence
AI summaryⓘ
The authors look at how people pick and mix datasets to train Vision Language Models (VLMs) and find that current methods are mostly based on guesswork. They propose a new system called DecoupleMix that breaks down data selection into two parts: deciding how much data to use from different classes and choosing the best datasets within those classes based on quality and difficulty. Their method uses math to find the best mix and helps decide what data to add next. Tests show their method works better than simple guessing and that good mixes found on small tests also work well at larger scales.
Vision Language Modelsdata curationpretrainingmixture optimizationinter-class ratiosintra-class ratiosdataset qualitydataset difficultyconvex optimizationmultimodal datasets
Authors
Jiahao Xie, Zhongbin Guo, Qianle Wang, Ruiqi Lu, Dongling Xiao, Wanxuan Sun, Cheng Yang
Abstract
While data curation for Vision Language Models (VLMs) is increasingly active, public practice for constructing pretraining mixtures remains largely heuristic: practitioners stack datasets that pass quality filters, set cross-domain ratios by intuition, and lack a principled, attributable criterion for admitting new data, while frontier recipes remain undisclosed. We formulate data construction as a systematic mixture-optimization problem and turn it into a reproducible engineering discipline by decoupling the mixture into two orthogonal sub-problems: inter-class ratios across capabilities and intra-class ratios within a category. For inter-class allocation, we use a single-variable iterative search; for intra-class composition, we apply a multidimensional, dataset-level assessment scoring Quality and Difficulty, and formulate selection as a constrained convex optimization with a diversity objective. The DecoupleMix framework delivers two critical capabilities: guiding what data to collect next and rendering dataset validation a controlled, attributable experiment. Experiments show our approach consistently surpasses heuristic baselines. Moreover, optimal ratios discovered on small-scale proxies transfer seamlessly to larger scales without retuning. Using 80B additional multimodal continue-pretraining tokens, our VLM is competitive with strong open-source models trained with substantially larger multimodal budgets.