已收录文章摘要，完整内容可前往原文阅读。
文章来源：ICLR 2026 论文题目：FlashVID: Efficient Video Large Language Models via Training-free Tree-based Spa