[人人能懂AI前沿] 从奖励骇客到深度推理，我们需要什么样的进化智慧?

在追求智能进化的路上，AI竟然也会像人一样陷入“KPI死胡同”、学会“钻空子”，甚至在层层伪装下变成只会背答案的“伪学霸”。本期节目，我们将拆解5篇最新论文，看看这些聪明的算法是如何通过“参考重置”、“反向回溯”与“逻辑解构”来完成进化的。这不仅是AI的成长课，更是我们每个人打破思维定式、拒绝原地打转的进阶指南。

00:00:29 为什么越想走捷径，反而越容易原地踏步，来自AI大模型训练的成长启示

00:05:42 哪怕是犯错，也别犯同样的错，关于AI如何“钻空子”的启示录

00:09:41 AI怎么学会思考，真正的聪明，都懂得“退一步”

00:13:45 如何识破人工智能的“伪学霸”？

00:17:20 你以为的因果，可能只是某种“继承”，从AI如何做算术说起

本期介绍的几篇论文：

[LG] A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets

[Harvard University & UIUC & University of California, San Diego]

https://arxiv.org/abs/2611.14467

---

[LG] Wrong the Same Way: Cross-Prompt Sharing Both Suppresses and Amplifies Reward Hacking in RLVR

[CMU & University of Washington]

https://arxiv.org/abs/2611.13218

---

[AI] Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth

[UC Berkeley]

https://arxiv.org/abs/2611.13594

---

[LG] Evaluating Compositional Generalization in Transformers: The Role of Composition Equivalence and Module Coverage

[University of Massachusetts Amherst & Google DeepMind]

https://arxiv.org/abs/2611.13498

---

[LG] Linear Probes to Bayesian Manifolds: Geometry of Arithmetic in Large Language Models

[CMU & Indian Institute of Technology Jodhpur & Indian Institute of Technology Roorkee]

https://arxiv.org/abs/2611.12998