![[人人能懂AI前沿] AI学霸的五张笔记：关于努力、谦逊、效率、选择与沟通](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
你有没有想过，最高效的学习，可能不是埋头苦干，而是学会“断舍离”？本期节目，我们将一起打开几篇最新论文，探讨AI如何向我们展示“聪明地努力”的全新境界。我们会看到，AI不仅开始筛选值得学习的“心动时刻”，还学会了在没把握时坦诚地说“我不知道”。更神奇的是，它们正通过“关键点教学法”和“性价比眼镜”，在复杂的任务中找到最高效的路径，并反思“会做题”与“会教题”的深刻区别。准备好了吗？让我们一起探索AI如何变得更精准、更谦逊、也更智慧！
00:41:25 别再无效努力了，学霸的秘诀是“断舍离”
00:06:25 那个“无所不知”的AI，为什么开始说“我不知道”了？
00:12:36 聪明地偷懒，AI训练的“性价比”之道
00:18:14 AI大模型选择困难症？这里有副“性价比”眼镜
00:23:47 “高手”的笔记，为什么你看不懂？
本期介绍的几篇论文：
[LG] Does This Gradient Spark Joy?
[Google DeepMind]
https://arxiv.org/abs/2603.20526
---
[LG] Causal Evidence that Language Models use Confidence to Drive Behavior
[Google DeepMind]
https://arxiv.org/abs/2603.22161
---
[LG] PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost
[NVIDIA & UC Berkeley]
https://arxiv.org/abs/2603.21383
---
[CL] Expected Reward Prediction, with Applications to Model Routing
[Stanford University & Google DeepMind]
https://arxiv.org/abs/2603.20217
---
[CL] Measuring Reasoning Trace Legibility: Can Those Who Understand Teach?
[CMU]
https://arxiv.org/abs/2603.20508
00:41:25 别再无效努力了，学霸的秘诀是“断舍离”
00:06:25 那个“无所不知”的AI，为什么开始说“我不知道”了？
00:12:36 聪明地偷懒，AI训练的“性价比”之道
00:18:14 AI大模型选择困难症？这里有副“性价比”眼镜
00:23:47 “高手”的笔记，为什么你看不懂？
本期介绍的几篇论文：
[LG] Does This Gradient Spark Joy?
[Google DeepMind]
https://arxiv.org/abs/2603.20526
---
[LG] Causal Evidence that Language Models use Confidence to Drive Behavior
[Google DeepMind]
https://arxiv.org/abs/2603.22161
---
[LG] PivotRL: High Accuracy Agentic Post-Training at Low Compute Cost
[NVIDIA & UC Berkeley]
https://arxiv.org/abs/2603.21383
---
[CL] Expected Reward Prediction, with Applications to Model Routing
[Stanford University & Google DeepMind]
https://arxiv.org/abs/2603.20217
---
[CL] Measuring Reasoning Trace Legibility: Can Those Who Understand Teach?
[CMU]
https://arxiv.org/abs/2603.20508
