你有没有想过，我们能像做微创手术一样，在AI思考的瞬间“拨乱反正”，引导它向善吗？或者，让昂贵的AI训练学会“温故知新”，把扔掉的经验变废为宝？本期节目，我们将一起探索几篇最新论文，看看科学家们如何教会AI遵守自己立下的规矩，如何让它既会“看路”又会“造景”，甚至，如何为它补上一堂生动的物理课，让它的想象力更符合现实。准备好了吗？让我们马上出发！

00:00:34 给AI的大脑装一个“概念导航”

00:06:53 AI训练的高效秘诀，好东西值得再用一次

00:12:07 如何看穿一个AI的“人设”？

00:16:56 AI新思路，想看清世界，先学会走路

00:23:07 为什么AI生成的视频总感觉“假”？答案藏在物理学里

本期介绍的几篇论文：

[LG] Dictionary-Aligned Concept Control for Safeguarding Multimodal LLMs  

[University of Pennsylvania & Amazon]  

https://arxiv.org/abs/2604.08846 

---

[LG] Efficient RL Training for LLMs with Experience Replay  

[FAIR at Meta]  

https://arxiv.org/abs/2604.08706 

---

[CL] Do LLMs Follow Their Own Rules? A Reflexive Audit of Self-Stated Safety Policies  

[Microsoft]  

https://arxiv.org/abs/2604.09189 

---

[CV] Rays as Pixels: Learning A Joint Distribution of Videos and Camera Trajectories  

[Meta AI]  

https://arxiv.org/abs/2604.09429 

---

[CV] PhysInOne: Visual Physics Learning and Reasoning in One Suite  

[vLAR Group]  

https://arxiv.org/abs/2604.09415