想让AI变聪明，一定要把它做得更大、训练得更久吗？最新论文告诉我们，恰恰相反！今天我们要聊聊，如何通过为AI“挖一口深井”来代替疯狂“堆料”，如何给它一颗“后悔药”让小模型也能超越巨无霸。我们还会揭秘AI训练的“快进键”是如何实现的，以及如何给AI一份“世界地图”和一位“高手陪练”，让它拥有永不枯竭的好奇心和可被预测的光明未来。

00:00:35 人工智能“开窍”的秘密，与其堆料，不如挖井

00:05:53 给机器一颗“后悔药”，小模型如何反超大模型？

00:12:02 AI训练的“快进键”，为什么只练开头就能猜到结尾？

00:17:12 AI“路痴”自救指南，如何拥有永不枯竭的好奇心

00:23:16 AI养成记，如何给一个“笨”模型当好陪练？

本期介绍的几篇论文：

[LG] Equilibrium Reasoners: Learning Attractors Enables Scalable Reasoning

[CMU]

https://arxiv.org/abs/2605.21488

---

[AI] Probabilistic Tiny Recursive Model

[Mila – Quebec AI Institute]

https://arxiv.org/abs/2605.19943

---

[LG] You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories

[University of Virginia]

https://arxiv.org/abs/2605.21468

---

[LG] Remember to be Curious: Episodic Context and Persistent Worlds for 3D Exploration

[University of Toronto & UC Berkeley & Wayve]

https://arxiv.org/abs/2605.22814

---

[CL] Forecasting Downstream Performance of LLMs With Proxy Metrics

[Mila – Quebec AI Institute & McGill University]

https://arxiv.org/abs/2605.18607