今天的几篇最新论文，将带你见证AI如何从“死记硬背的做题家”蜕变成“讲究体面的真正高手”。我们首先会看到AI如何在严苛的防作弊机制下逼出真实的自我进化，又如何向人类借取经验技能包、学会做事讲规矩；紧接着，两篇硬核的最新论文将展示最聪明的减法：通过外挂记忆字典给昂贵算力降载，以及仅用1%的高信噪比反馈超越100%的全量穷忙；最后，我们还将解锁一套不需要窥探个人隐私、仅凭宏观数据就能精准预判群体未来走向的动力学模型，为你奉上一场前沿技术与认知进阶的双重盛宴！00:00:44 摆脱“题海战术”，人工智能教给普通人的自我进化法则00:06:07 别只教AI“做对”，还要教它“讲究”00:10:57 聪明人的“算计”，从AI学会恰当偷懒，看我们如何省下最贵的心智成本00:15:22 为什么1%的努力胜过100%的穷忙？从AI的“极简学习法”说起00:20:00 捕捉水流的形状，我们如何预测一个群体的未来？本期介绍的几篇论文：[LG] MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement[LLM-Core Xiaomi]https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL/blob/main/MiMo_V2_6_technical_report.pdf---[LG] Reinforcing Agents with Collective Skills[NVIDIA]https://github.com/NVlabs/Skill2Env/blob/main/paper/Skill2Env_arXiv.pdf---[LG] Memory Attention[J Kang]https://arxiv.org/abs/2609.28399---[LG] 1% of Tokens Can Be Enough: On Gradient Estimation in On-Policy Distillation[H Sheng, Z Ye, H Wang, J Wang… (MBZUAI & Ant Group)]https://arxiv.org/abs/2609.24432---[LG] Learning Collective Dynamics with Differentiable Gaussian Representations[J Ma, M Zhang, X Yang, Y Gao… (OranAI & Northeastern University)]https://arxiv.org/abs/2609.28405在小宇宙查看该单集文稿

Ask this episode

“What did the council decide on item 42?” — answered with the speaker and the moment.

Transcript

Related