![[人人能懂AI前沿] AI的元认知革命：从自信校准、演化微调到偏好重对齐](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
你有没有想过，AI的“内心世界”是什么样的？本期我们要聊的几篇最新论文，就像是为我们打开了AI心智的几扇窗：当AI说“我很确定”时，它可能只是下定了决心；而一个“无欲无求”的旁观者AI，或许才是通往安全的新路径。我们还会看到，AI如何通过“开窍”学会跨界创新，如何用“错题本”学会自我反思，以及我们普通人如何拥有一本不用编程的“AI调校手册”。准备好了吗？让我们一起潜入AI思考的深处。
00:00:35 AI说“我确定”的时候，它到底在确定什么？
00:08:51 AI进化新思路，当个“旁观者”，而不是“操盘手”
00:15:50 让聪明的模型，学会“开窍”
00:20:02 一个会反思的AI，如何从犯错中学会正确答案
00:24:44 驯服AI，一个不用编程的调校手册
本期介绍的几篇论文：
[LG] Reported Confidence in LLMs Tracks Commitment More Than Correctness
[Google DeepMind]
https://arxiv.org/abs/2606.29490
---
[AI] Safety from Honesty in a Disinterested AI Predictor
[LawZero & Arb Research]
https://arxiv.org/abs/2606.29657
---
[CL] Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks
[University of Minnesota & CMU & KAIST]
https://arxiv.org/abs/2606.29082
---
[AI] Flow Reasoning Models: Scaling Reasoning Through Iterative Self-Refinement
[Georgia Tech & MIT]
https://arxiv.org/abs/2606.29150
---
[CL] REAR: Test-time Preference Realignment through Reward Decomposition
[Nanyang Technological University & UC Berkeley]
https://arxiv.org/abs/2606.30339
00:00:35 AI说“我确定”的时候，它到底在确定什么？
00:08:51 AI进化新思路，当个“旁观者”，而不是“操盘手”
00:15:50 让聪明的模型，学会“开窍”
00:20:02 一个会反思的AI，如何从犯错中学会正确答案
00:24:44 驯服AI，一个不用编程的调校手册
本期介绍的几篇论文：
[LG] Reported Confidence in LLMs Tracks Commitment More Than Correctness
[Google DeepMind]
https://arxiv.org/abs/2606.29490
---
[AI] Safety from Honesty in a Disinterested AI Predictor
[LawZero & Arb Research]
https://arxiv.org/abs/2606.29657
---
[CL] Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks
[University of Minnesota & CMU & KAIST]
https://arxiv.org/abs/2606.29082
---
[AI] Flow Reasoning Models: Scaling Reasoning Through Iterative Self-Refinement
[Georgia Tech & MIT]
https://arxiv.org/abs/2606.29150
---
[CL] REAR: Test-time Preference Realignment through Reward Decomposition
[Nanyang Technological University & UC Berkeley]
https://arxiv.org/abs/2606.30339
外观
