💡 This paper presents a systematic study on long-context continued pre-training for vision-language models, deriving key training recipes (e.g., long-doc VQA data, balanced length distribution, retrieva