你敢相信吗？给大模型“戴上眼罩”隔离全局信息，或是把海量长文本优雅地压缩成“背景底噪”，竟然能让推理泛化更稳、处理速度翻倍。今天这期节目要介绍的五篇最新论文彻底颠覆了直觉：不仅有五个靠“可验证看板”接力逆袭33个单兵天才的AI协作网络，更有用形式化证明扯下“代码测试满分”遮羞布的严苛审计。我们还将看到，只需给调度员开一扇“回看历史目光”的小窗，模型就能学会知错就收的从容。带上你的好奇心，让我们一起潜入这五项最新论文构建的全新认知世界！

00:00:43 给大模型戴上一副“马眼罩”，为什么知道得越少，反而算得越准？

00:06:09 聪明的注意力，从来不是简单的一刀切

00:12:48 为什么五个会聊天的AI，能打败三十三个单打独斗的天才？

00:18:52 别被“测试通过”骗了，当AI学会自我证明，它真正的死穴在哪里？

00:24:40 别让调度员“蒙着眼睛派活”，大模型悄悄变聪明的隐秘回路

本期介绍的几篇论文：

[CL] Recursive Language Models Generalize Out of Domain

[Toyota Technological Institute at Chicago]

https://arxiv.org/abs/2609.2083

---

[LG] Elastic Threshold Attention：Learned Contextual Sparsity for Long-Context Decoding

[Google]

https://arxiv.org/abs/2609.20888

---

[LG] Scaling Discovery through Test-Time Communication

[UC Berkeley & Microsoft Research]

https://arxiv.org/abs/2609.21032

---

[LG] SWE-Proof：Can Language Models Resolve Real-World Issues with Machine-Checked Proofs?

[UC Berkeley & Georgia Tech & UIUC]

https://arxiv.org/abs/2609.21190

---

[AI] Attention-Aware Routing: Coupling Routing and Attention in MoEs

[National Technical University of Athens & University of Bern]

https://arxiv.org/abs/2609.20974