![[人人能懂AI前沿] AI的成长三部曲：学会约束、学会思考、学会记忆](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
如果AI像个学生，我们该如何教育它？本期节目，我们将一起探索几篇最新论文带来的惊人答案：我们将看到，如何用一根充满智慧的“弹力绳”防止AI“学疯了”；如何用一棵“假设树”教会AI像科学家一样累积经验；我们还会举办一场AI记忆力大赛，看看究竟是“死记硬背”还是“内在结构”更胜一筹；最后，我们将揭示一种让AI“开卷的我”去教“闭卷的我”的神奇训练法，并学会如何像外科医生一样，为AI精准“切除”坏习惯。准备好了吗？让我们一起看看，人类是如何教会AI“学习如何学习”的。
00:00:42 AI“学疯了”怎么办？一根“弹力绳”的智慧
00:06:42 如何让AI像科学家一样思考？
00:13:26 你的记忆，是“看过”还是“记住”了？
00:19:22 AI训练的新思路，优等生是如何“开卷”带“闭卷”的？
00:25:10 我们给AI的“好评”，正在让它变“笨”吗？
本期介绍的几篇论文：
[LG] Rethinking the Divergence Regularization in LLM RL
[Tencent Hunyuan & NUS]
https://arxiv.org/abs/2606.09821
---
[CL] Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
[Microsoft Research & Renmin University of China]
https://arxiv.org/abs/2606.11926
---
[CV] Echo-Memory: A Controlled Study of Memory in Action World Models
[The University of Hong Kong & Joy Future Academy, JD & The Chinese University of Hong Kong]
https://arxiv.org/abs/2606.09803
---
[LG] Rubric-Guided Self-Distillation: Post-Training Without Rubric Verifiers
[Scale AI]
https://arxiv.org/abs/2606.12507
---
[LG] Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal
[GOODFIRE]
https://arxiv.org/abs/2606.12360
00:00:42 AI“学疯了”怎么办？一根“弹力绳”的智慧
00:06:42 如何让AI像科学家一样思考？
00:13:26 你的记忆，是“看过”还是“记住”了？
00:19:22 AI训练的新思路，优等生是如何“开卷”带“闭卷”的？
00:25:10 我们给AI的“好评”，正在让它变“笨”吗？
本期介绍的几篇论文：
[LG] Rethinking the Divergence Regularization in LLM RL
[Tencent Hunyuan & NUS]
https://arxiv.org/abs/2606.09821
---
[CL] Toward Generalist Autonomous Research via Hypothesis-Tree Refinement
[Microsoft Research & Renmin University of China]
https://arxiv.org/abs/2606.11926
---
[CV] Echo-Memory: A Controlled Study of Memory in Action World Models
[The University of Hong Kong & Joy Future Academy, JD & The Chinese University of Hong Kong]
https://arxiv.org/abs/2606.09803
---
[LG] Rubric-Guided Self-Distillation: Post-Training Without Rubric Verifiers
[Scale AI]
https://arxiv.org/abs/2606.12507
---
[LG] Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal
[GOODFIRE]
https://arxiv.org/abs/2606.12360
外观
