Multimodal models: better together, cheaper than expected, and apparently vision is lazy

This paper systematically investigates how vision and language modalities interact during unified multimodal pretraining, identifying four key findings: knowledge flows asymmetrically across modalities, data complexity determines synergy vs. competition, early unification outperforms late alignment (avoiding a 'vision laziness' problem), and efficient recipes can match strong generative performance at 5% compute cost. The authors validate their insights by training multiple 13.5B parameter MoE models on 2T tokens.