![[人人能懂AI前沿] 从“差值学习”到“思考快车道”：AI智能升级的五条新路径](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
你有没有想过，为什么机器人总是记不住自己干了什么？AI的“老师”会不会偷偷从网页评论区学习知识？今天，我们就来聊聊几篇最新论文带来的奇妙启发：我们将一起探索如何给机器人装上“好记性”，如何揪出AI知识里的“隐藏毒药”，并揭示让AI学会高效“差值学习”、拥有“一步到位”想象力，甚至打通“思考快车道”的秘密。
00:00:29 机器人笨手笨脚？可能只是记性不好
00:05:39 你的AI老师，可能正在偷看网页评论区
00:13:11 高手精进的秘密，不止是模仿，更是学“差值”
00:18:07 让机器人学会“一步到位”的想象力
00:23:37 AI思考的“快车道”
本期介绍的几篇论文：
[RO] RoboTTT: Context Scaling for Robot Policies
[NVIDIA]
https://arxiv.org/abs/2607.15275
---
[CL] Pretraining Data Can Be Poisoned through Computational Propaganda
[University of Washington]
https://arxiv.org/abs/2607.15267
---
[LG] On-Policy Delta Distillation
[NAVER AI Lab]
https://arxiv.org/abs/2607.15161
---
[RO] DriftWorld: Fast World Modeling through Drifting
[MIT & Harvard University]
https://arxiv.org/abs/2607.15065
---
[CL] T²MLR: Transformer with Temporal Middle-Layer Recurrence
[Princeton University]
https://arxiv.org/abs/2607.15178
00:00:29 机器人笨手笨脚？可能只是记性不好
00:05:39 你的AI老师，可能正在偷看网页评论区
00:13:11 高手精进的秘密，不止是模仿，更是学“差值”
00:18:07 让机器人学会“一步到位”的想象力
00:23:37 AI思考的“快车道”
本期介绍的几篇论文：
[RO] RoboTTT: Context Scaling for Robot Policies
[NVIDIA]
https://arxiv.org/abs/2607.15275
---
[CL] Pretraining Data Can Be Poisoned through Computational Propaganda
[University of Washington]
https://arxiv.org/abs/2607.15267
---
[LG] On-Policy Delta Distillation
[NAVER AI Lab]
https://arxiv.org/abs/2607.15161
---
[RO] DriftWorld: Fast World Modeling through Drifting
[MIT & Harvard University]
https://arxiv.org/abs/2607.15065
---
[CL] T²MLR: Transformer with Temporal Middle-Layer Recurrence
[Princeton University]
https://arxiv.org/abs/2607.15178
外观
