multimodalfactneutralTraining data quality is more important than model scale for VLM robustness, with curated datasets yielding up to 25% improvements over uncurated alternatives at comparable scaleComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28211v1