你有没有想过，AI是在帮你分析，还是在高级地“说服”你？我们总希望AI像个完美的老师，但如果它只会给标准答案，甚至连老师的偏见都一并继承，那会怎样？而为了让AI学得更好，我们不仅要为它的“记忆”做体检，甚至还要教会它一项人类的高级智慧：学会放弃。今天，我们就从五篇最新的论文出发，看看AI是如何在说服、学习和思考的边界上，进行着一场静悄悄的认知革命。

00:00:33 当AI学会了“高级说服”，你的大脑还够用吗？

00:06:00 如何给AI做一次“记忆体检”？

00:12:34 AI只会“标准答案”？那可就危险了

00:18:04 高手过招，如何避免被师傅“带偏”？

00:23:19 训练AI的真谛，学会放弃，才能得到更多

本期介绍的几篇论文：

[AI] Evaluating Language Models for Harmful Manipulation  

[Google DeepMind & Google]  

https://arxiv.org/abs/2603.25326 

---

[CL] Estimating near-verbatim extraction risk in language models with decoding-constrained beam search  

[Stanford & Cornell]  

https://arxiv.org/abs/2603.24917 

---

[LG] Reaching Beyond the Mode: RL for Distributional Reasoning in Language Models  

[MIT]  

https://arxiv.org/abs/2603.24844 

---

[LG] Residual-as-Teacher: Mitigating Bias Propagation in Student--Teacher Estimation  

[MIT]  

https://arxiv.org/abs/2603.25466 

---

[CL] Prune as You Generate: Online Rollout Pruning for Faster and Better RLVR  

[University of Illinois at Urbana-Champaign]  

https://arxiv.org/abs/2603.24840