你是否想过，如何让“口是心非”的AI学会言行一致，又如何让手机App在“懂你”的同时做到“不认识你”？本期节目，我们将一起揭秘几篇最新论文，看看科学家们如何用“左右互搏”大法驯服AI，用“精准激励”破解AI的“中年危机”，甚至将AI的“直觉”直接翻译成我们能读懂的代码。准备好了吗？让我们一起出发！

00:00:27 驯服AI野马，从“口是心非”到“知行合一”

00:06:49 鱼与熊掌，如何让App既“懂你”又“不认识你”？

00:11:27 如何破解AI训练的“中年危机”？

00:16:48 让机器人学会“看样学样”，总共分几步？

00:22:52 把AI的“直觉”翻译成代码，会发生什么？

本期介绍的几篇论文：

[LG] Self-CTRL: Self-Consistency Training with Reinforcement Learning

[MIT CSAIL]

https://arxiv.org/abs/2606.18327

---

[LG] Private Learning with Public Feature Conditioning

[AWS Agentic AI & Microsoft & Google Research]

https://arxiv.org/abs/2606.18773

---

[LG] STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability

[Tencent Hunyuan & Tsinghua University]

https://arxiv.org/abs/2606.19236

---

[RO] Do as I Do: Dexterous Manipulation Data from Everyday Human Videos

[UC Berkeley]

https://arxiv.org/abs/2606.19333

---

[LG] Explaining Attention with Program Synthesis

[NJIT & MIT EECS]

https://arxiv.org/abs/2606.19317