你有没有想过，我们该如何教会AI那些没有标准答案的事？本期节目，我们将一起探讨几篇最新论文带来的奇妙思路：从把AI的“裁判”换成“教练”，到给机器人换上一副“高清眼镜”，再到为AI装上一个能自我更新的“智能错题本”；我们甚至会发现，让AI在“梦境”里胡思乱想，以及在它钻牛角尖时悄悄“推”它一把，或许才是通往更强人工智能的捷径。

00:00:30 AI进化论，别当裁判，请当教练

00:05:33 让机器人更灵巧，不一定要给它一个更大的脑子

00:10:51 如何给AI装上一个“智能错题本”？

00:16:49 你的大脑不是硬盘，而是一座创意的梦工厂

00:22:02 给AI装个导航，让它少走冤枉路

本期介绍的几篇论文：

[LG] LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks

[Microsoft Research]

https://arxiv.org/abs/2607.18110

---

[RO] Patch Policy: Efficient Embodied Control via Dense Visual Representations

[New York University]

https://arxiv.org/abs/2607.18236

---

[AI] Fantastic Adaptive Taxonomies and How to Use Them

[UC Berkeley]

https://arxiv.org/abs/2607.16387

---

[LG] Discovery by Dreaming: Cross-Domain Recombination in Artificial Memory

[University of Chicago & Stanford University]

https://arxiv.org/abs/2607.16256

---

[LG] Can We Break LLMs Out of Self-Loops? Fine-Grained Reasoning Control with Activation Steering

[UC San Diego & Adobe Research]

https://arxiv.org/abs/2607.18100