EVENT DOSSIER
The research team proposed VESTA, a trained-free long-video intelligent agent system. This system infers focus, recall, or contrast retrieval strategies and evidence account configurations during the exploration phase through an intention router, and employs a routing-based - verification - consolidation cycle for task processing. VESTA uses multimodal evidence manipulation to transform retrieval results into observations, and integrates location, source, conflicts, and verification results through a time evidence ledger to form an adaptive view to guide subsequent actions. In the Video-MME-v2 test, its average accuracy improved by 2.7 points compared to VideoARM; in the LongVideoBench long-subset and LVBench, it improved by 6.9 points and 1.5 points respectively, and performed equally with VideoARM in EgoSchema.
提出 VESTA，一种无需训练的长视频智能体，通过意图路由器在探索前推断聚焦、召回或对比检索策略及证据账户配置。该智能体采用路由条件获取 - 验证 - 巩固循环，利用多模态证据操作将检索结果转化为观察值，并由时间证据账本整合为包含位置、来源、冲突及验证结果的自适应视图以指导后续行动。在 Video-MME-v2 上，VESTA 平均准确率较 VideoARM 提升 2.7 分；在 LongVideoBench 长子集和 LVBench 上分别提升 6.9 分和 1.5 分，并在 EgoSchema 上与 VideoARM 持平。