少用 43% 的 Token，保留 98.9% 的质量：VisionWeave 如何像花钱一样分配视觉 Token

43% Fewer Tokens， 98.9% of the Quality: How VisionWeave Spends Visual Tokens Like Money

VisionWeave, a method from Alibaba's Qwen team, reduces visual token usage by 43% while maintaining 98.9% of model performance, achieving this by adaptively allocating tokens based on image content rather than using a fixed budget.