00:00:37 你的AI管家，靠谱吗？一份来自未来的安全报告

00:04:40 AI“发疯”？科学家找到了它的“性格开关” 

00:09:33 比结果更重要的，是“想明白”的过程 

00:14:09 AI的“降维打击”：复杂世界里的简单活法 

00:18:23 AI的“暖男”人设，可能是个陷阱？

本期介绍的几篇论文：

[LG] Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition  

[Gray Swan AI]  

https://arxiv.org/abs/2507.20526  

---

[CL] Persona Vectors: Monitoring and Controlling Character Traits in Language Models  

[Anthropic Fellows Program & Constellation]  

https://arxiv.org/abs/2507.21509  

---

[LG] RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents  

[Tencent]  

https://arxiv.org/abs/2507.22844  

---

[LG] Geometry of Neural Reinforcement Learning in Continuous State and Action Spaces  

[Brown University & Amazon Web Services]  

https://arxiv.org/abs/2507.20853  

---

[CL] Training language models to be warm and empathetic makes them less reliable and more sycophantic  

[University of Oxford]  

https://arxiv.org/abs/2507.21919  

---

[CL] On The Role of Pretrained Language Models in General-Purpose Text Embeddings: A Survey  

[Not explicitly stated, survey paper]  

https://arxiv.org/abs/2507.20783