Scaling Native Multimodal Pre-Training From Scratch

  • 类型：arxiv
  • 标识：2607.22043
  • 链接：https://arxiv.org/abs/2607.22043
  • 主分类：multimodal
  • 形态：method
  • 被引：1
  • 被引来源：Semantic Scholar
  • S2被引：1
  • 影响力被引：0
  • TLDR：This empirical research establishes the essential groundwork for predictably scaling multimodal foundation models by modeling the influence of data composition on compute laws and allocation exponents and derive an efficiency frontier specifying precise configurations of model size, token count, and data mixture.
  • 副分类：engineering
  • 待LLM分类：否
  • 标题中文：Scaling Native Multimodal Pre-Training From Scratch
  • TLDR中文：本实证研究通过建模数据组成对计算定律及分配指数的影响，推导出指定模型规模、token 数和数据混合精确配置的效率前沿，为可预测地扩展多模态基础模型奠定了必要基础。
  • 来源文件：
  • /inbox/tom/_candidates/2026-07-27-agent-rag-longcontext-candidates.json
  • [S2 enrich]