视频选题榜 2026-08-06

• https://arxiv.org/abs/2608.03979 · Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent · 解耦 perception-exploration pipeline + stage-wise tool unlocking 强制先看视频再上网，35B-A3B 64.0% 超 Claude-4.5-Sonnet 59.0% 5.0 points · round=R1- https://arxiv.org/abs/2608.02218 · PosterMELD: Multi-Agent Paper-to-Poster Generation for Controllable Design Diversity with Editable Print-Ready Outputs · 模板条件化 + 容量感知槽位 + 确定性门控/VLM 审核把论文转海报做成失败可路由的多 Agent 流水线 · round=R2
• https://arxiv.org/abs/2608.01127 · MiniWorld: Democratizing the Training of Video World Models from Scratch · 流式视频世界模型民主化训练配方 · block-causal Video DiT + chunk-wise noise + rolling KV + 异步去噪 + 单 8-GPU 数天内训完 · round=R3