本期我们要聊一个核心问题：我们总觉得AI是个神秘的黑箱，但最新的研究正在像做“脑部扫描”一样，试图撬开它。我们将看到，一个“满分或零分”的简单规则，就能教会AI诚实；又如何派出一个“AI侦探”，揪出潜伏的恶意模型。接着，我们会深入AI的“思考过程”，看看聪明的“大脑”和聪明的“搜索引擎”哪个更重要，以及如何让AI通过“犯错”来演化出正确答案，甚至把它的复杂推理拆解成一个个可以遥控的“思想积木”。准备好了吗？让我们一起深入AI的内心世界。

00:00:41 AI的“不说谎”训练：满分或零分

00:05:29 AI界的“无间道”：如何揪出披着羊皮的狼？

00:10:39 聪明的大脑，和聪明的搜索引擎，哪个更重要？

00:16:14 犯错没关系，只要你“改对”的概率比“改错”大一点点

00:21:22 拆解AI大脑：它思考时在想什么？

本期介绍的几篇论文：

[CL] Train for Truth，Keep the Skills：Binary Retrieval-Augmented Reward Mitigates Hallucinations

[University of Washington & Allen Institute for AI (Ai2)]

https://arxiv.org/abs/2510.17733

---

[LG] Detecting Adversarial Fine-tuning with Auditing Agents

[Anthropic]

https://arxiv.org/abs/2510.16255

---

[LG] Prior Makes It Possible：From Sublinear Graph Algorithms to LLM Test-Time Methods

[Toyota Technological Institute at Chicago & Columbia University & Google Research]

https://arxiv.org/abs/2510.16609

---

[CL] Deep Self-Evolving Reasoning

[Microsoft Research Asia & Peking University]

https://arxiv.org/abs/2510.17498

---

[LG] Algorithmic Primitives and Compositional Geometry of Reasoning in Language Models

[Columbia University & University of California Los Angeles & Harvey Mudd College]

https://arxiv.org/abs/2510.15987

[人人能懂] 从二元奖惩、动态审计到思维积木

AI可可AI生活