DecoupleMix turns VLM data recipes into a reproducible optimization problem

Jiahao Xie · hf · 2026-07-28

DecoupleMix reframes VLM pretraining data curation as a mixture-optimization problem rather than a heuristic stacking exercise.
It decouples data recipe design into two orthogonal parts: inter-class ratios across capabilities and intra-class ratios within a category.
Inter-class allocation is solved with a single-variable iterative search, while intra-class selection uses quality/difficulty scoring plus constrained convex optimization with a diversity objective.
The framework is meant to make dataset validation reproducible and attributable; experiments show it beats heuristic baselines, and ratios found on small proxies transfer to larger scales without retuning.

More from Multimodal