![[人人能懂AI前沿] 从倒置沙漏、导师容错，到历史错题与抽象穿梭](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
今天我们要拆解的五篇最新论文，正在彻底颠覆关于大模型与效率的默认规则：从打破行业惯例、把模型变窄的“沙漏架构”，到允许瑕疵反而超越上限的“导师制解码”；从靠翻看“历史错题本”避免原地打转的黑盒学习，到自我修剪冗余的“递归进化”，以及在宏观与底层之间自由穿梭的“抽象阶梯”。这不仅是一次硬核技术的减负提速之旅，更是一份藏在代码算法里、人人皆可借用的自我迭代指南。
00:00:35 别被“行业惯例”限制了想象，把沙漏倒过来，世界就顺畅了
00:04:46 放下对完美的死磕，为什么“允许瑕疵”反而能成就更好？
00:09:48 让你突飞猛进的秘密，往往藏在被遗忘的“旧账”里
00:13:41 如何打破成长的天花板？这篇AI前沿论文藏着一套“自我进化”的破局心法
00:18:05 做人做事，要学会在“抽象的阶梯”上自由上下
本期介绍的几篇论文：
[CL] Revisiting the Shape Convention of Transformer Language Models
[MediaTek Research]
https://arxiv.org/abs/2602.06471
---
[LG] Mentored Decoding: Faster Inference meets Boosting
[Google]
https://arxiv.org/abs/2609.30474
---
[CL] Persistent Negatives for Adversarial Black-Box On-Policy Distillation
[Meta AI]
https://arxiv.org/abs/2609.3086
---
[CL] Recursive Self-Improvement via On-Policy Distillation for Reasoning
[Meta AI]
https://arxiv.org/abs/2609.3065
---
[AI] Up and Down the Abstraction Ladder: Code-Based Skills for Language Agents
[University of Warsaw & Princeton University & IDEAS NCBR]
https://arxiv.org/abs/2609.3107
00:00:35 别被“行业惯例”限制了想象，把沙漏倒过来，世界就顺畅了
00:04:46 放下对完美的死磕，为什么“允许瑕疵”反而能成就更好？
00:09:48 让你突飞猛进的秘密，往往藏在被遗忘的“旧账”里
00:13:41 如何打破成长的天花板？这篇AI前沿论文藏着一套“自我进化”的破局心法
00:18:05 做人做事，要学会在“抽象的阶梯”上自由上下
本期介绍的几篇论文：
[CL] Revisiting the Shape Convention of Transformer Language Models
[MediaTek Research]
https://arxiv.org/abs/2602.06471
---
[LG] Mentored Decoding: Faster Inference meets Boosting
[Google]
https://arxiv.org/abs/2609.30474
---
[CL] Persistent Negatives for Adversarial Black-Box On-Policy Distillation
[Meta AI]
https://arxiv.org/abs/2609.3086
---
[CL] Recursive Self-Improvement via On-Policy Distillation for Reasoning
[Meta AI]
https://arxiv.org/abs/2609.3065
---
[AI] Up and Down the Abstraction Ladder: Code-Based Skills for Language Agents
[University of Warsaw & Princeton University & IDEAS NCBR]
https://arxiv.org/abs/2609.3107
外观
