[人人能懂] 回音室、水印破解与选择之道

00:00:36 AI的“回音室”危机：当复印件开始复印自己

00:05:30 AI的“超级大脑”里，谁在干活，谁在摸鱼？

00:10:23 AI的“隐形纹身”：一张纸巾就能擦掉？

00:15:27 AI的“选择之道”：为何“条条大路通罗马”本身就是一种智慧？

00:20:39 AI的“开窍”秘诀：好学生为何不能太“自信”？

本期介绍的几篇论文：

[LG] ForTIFAI: Fending Off Recursive Training Induced Failure for AI Models  

[UC San Diego & Stanford University]  

https://arxiv.org/abs/2509.08972  

---

[CL] Steering MoE LLMs via Expert (De)Activation  

[University of California, Los Angeles & LMU Munich & Adobe Research]  

https://arxiv.org/abs/2509.09660  

---

[LG] Character-Level Perturbations Disrupt LLM Watermarks  

[University of Technology Sydney & Griffith University]  

https://arxiv.org/abs/2509.09112  

---

[LG] An entropy formula for the Deep Linear Network  

[G Menon, T Yu]  

https://arxiv.org/abs/2509.09088  

---

[CL] CDE: Curiosity-Driven Exploration for Efficient Reinforcement Learning in Large Language Models  

[Tencent AI Lab]  

https://arxiv.org/abs/2509.09675