2026.04.28 | 强化学习逼出几何一致视频；AI公司乐高式组队降本提效

【目录】本期的 15 篇论文如下：[00:24] 🌍 World-R1: Reinforcing 3D Constraints for Text-to-Video Generation（世界-R1：通过强化学习为文本到视频生成注入3D约束）[01:29] 🏢 From Skills to Talent: Organising Heterogeneous Agents as a Real-World Company（从技能到人才：将异构智能体组织为现实世界公司）[02:26] 🧠 ReVSI: Rebuilding Visual Spatial Intelligence Evaluation for Accurate Assessment of VLM 3D Reasoning（重建视觉空间智能评估：精准评估VLM三维推理能力）[03:23] 🛡 Vision-Language-Action Safety: Threats, Challenges, Evaluations, and Mechanisms（视觉-语言-动作安全：威胁、挑战、评估与机制）[04:12] 🖼 Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation（Tuna-2：像素嵌入在多模态理解与生成中击败视觉编码器）[05:02] 🤖 ClawMark: A Living-World Benchmark for Multi-Turn, Multi-Day, Multimodal Coworker Agents（ClawMark：面向多轮、多日、多模态协作者智能体的现实世界基准测试）[06:20] ✍ SketchVLM: Vision language models can annotate images to explain thoughts and guide users（SketchVLM：视觉语言模型可以通过图像标注来解释思维并引导用户）[07:17] 🔬 Rewarding the Scientific Process: Process-Level Reward Modeling for Agentic Data Analysis（奖励科学过程：面向智能体数据分析的过程级奖励建模）[08:24] ⚖ Taming Actor-Observer Asymmetry in Agents via Dialectical Alignment（通过辩证对齐驯服智能体中的行动者-观察者不对称性）[09:20] 🔀 Efficient Agent Evaluation via Diversity-Guided User Simulation（通过多样性引导的用户模拟实现高效智能体评估）[10:02] ⚡ For-Value: Efficient Forward-Only Data Valuation for finetuning LLMs and VLMs（For-Value：面向微调大语言模型和视觉语言模型的高效前向数据估值方法）[11:04] 🎬 OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer（全镜头剪切：基于镜头查询Transformer的整体关系型镜头边界检测）[12:03] 📷 UniGeo: Unifying Geometric Guidance for Camera-Controllable Image Editing via Video Models（UniGeo：通过视频模型实现相机可控图像编辑的统一几何引导）[12:49] 📄 TexOCR: Advancing Document OCR Models for Compilable Page-to-LaTeX Reconstruction（TexOCR：面向可编译页面到LaTeX重建的文档OCR模型进展）[13:56] 🔄 How Much Is One Recurrence Worth? Iso-Depth Scaling Laws for Looped Language Models（一次循环值多少？循环语言模型的等深度缩放定律）【关注我们】您还可以在以下平台找到我们，获得播客内容以外更多信息小红书: AI速递在小宇宙查看该单集文稿