MixAtlas: Uncertainty-aware Data Mixture Optimization for Multimodal LLM Midtraining

This paper was accepted at the Workshop on Navigating and Addressing Data Problems for Foundation Models (NADPFM) at ICLR 2026.

Principled domain reweighting can substantially improve sample efficiency and downstream generalization; however, data-mixture optimization for multimodal pretraining remains underexplored. Current multimodal training recipes tune mixtures from only a single perspective such as data format or task type. We introduce MixAtlas, a principled framework for compute-efficient multimodal mixture optimization via systematic domain decomposition and smaller proxy models...