[人人能懂] AI训练2.0：从“学霸”到“思考伙伴”的进化之路

00:00:26 AI的自我修养：当机器开始给自己出题

00:03:54 给AI装上一个会“犯嘀咕”的大脑 

00:08:57 你的AI客服，为什么总被人带跑偏？ 

00:13:36 AI的“速成班”：一个简单动作，让机器秒懂人心 

00:17:36 如何让你身边的AI“开窍”？ 

本期介绍的几篇论文：

[LG] Self-Questioning Language Models  

[CMU]  

https://arxiv.org/abs/2508.03682  

---

[LG] Cognitive Loop via In-Situ Optimization: Self-Adaptive Reasoning for Science  

[Microsoft]  

https://arxiv.org/abs/2508.02789  

---

[CL] Highlight & Summarize: RAG without the jailbreaks  

[Microsoft]  

https://arxiv.org/abs/2508.02872  

---

[CL] Cropping outperforms dropout as an augmentation strategy for training self-supervised text embeddings  

[University of Tübingen]  

https://arxiv.org/abs/2508.03453  

---

[LG] Agent Lightning: Train ANY AI Agents with Reinforcement Learning  

[Microsoft Research]  

https://arxiv.org/abs/2508.03680