DecoupleMix: Decoupled Ratio Search and Convex Allocation for Scalable VLM Data Recipes

Researchers introduced DecoupleMix to optimize vision-language model data mixtures by separating inter-class and intra-class ratio search.

The framework uses search, quality scoring, and convex allocation instead of hand-tuned heuristics to build scalable pretraining recipes.

DecoupleMix aims to make multimodal data curation more reproducible and transferable from small proxy runs to larger training setups.

The approach is designed to improve data efficiency and generalize ratio choices across continue-pretraining and large-scale VLM training.