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.

Original post →

More from Multimodal

• Google Gemini Omni Flash is used to make short videos at about 90 credits each — iamaliveix · 2026-07-29
• Microsoft removes Mage Flow and points to a more efficient Qwen-VL-class encoder — Dante_77A · 2026-07-29
• Opus 5 users are building full games, FPS prototypes, and 3D scenes in one shot — eyishazyer · 2026-07-29
• Seedance 2.0 prompt turns text-to-video into a chaotic bodycam street scene — techhalla · 2026-07-29
• A burger-cooking AI demo splits viewers between detail accuracy and ad-like polish — victor_explore · 2026-07-29
• SCAIL 2 survives character swaps, object permanence, and physics tests better than expected — blackmixture · 2026-07-29

Multimodal channel →