[人人能懂] 从团体赛、免费午餐到优雅忘记

想不想知道，如何通过一场“团体赛”规则，治好AI“偏科”的毛病？又如何不花一分钱，只靠一个古老的“背包问题”，就让AI训练效率飙升40%？甚至，我们将看到一个天生“失明”的AI，如何在想象力测试中击败人类。本期节目，我们将从最新的几篇论文出发，揭示AI如何学会团队合作、聪明用钱，甚至掌握了“蒙眼思考”和“断舍离”的智慧。

00:00:32 AI训练指南：如何让你的“学霸”不偏科？

00:05:50 AI训练的“免费午餐”：如何不花一分钱，让模型变得更聪明？

00:11:57 从“一指禅”到“组合拳”：AI学习的升级之路

00:16:54 AI的“心灵盲区”：它看不见，却想得更明白？

00:21:19 AI的记忆难题：如何优雅地“断舍离”？

本期介绍的几篇论文：

[LG] Polychromic Objectives for Reinforcement Learning  

[Stanford University]  

https://arxiv.org/abs/2509.25424 

---

[LG] Knapsack RL: Unlocking Exploration of LLMs via Optimizing Budget Allocation  

[ByteDance Seed & The Chinese University of Hong Kong]  

https://arxiv.org/abs/2509.25849 

---

[LG] Learning to Reason as Action Abstractions with Scalable Mid-Training RL  

[Apple]  

https://arxiv.org/abs/2509.25810 

---

[LG] Artificial Phantasia: Evidence for Propositional Reasoning-Based Mental Imagery in Large Language Models  

[Northeastern University]  

https://arxiv.org/abs/2509.23108 

---

[LG] Expected Attention: KV Cache Compression by Estimating Attention from Future Queries Distribution  

[NVIDIA & Sapienza University of Rome]  

https://arxiv.org/abs/2510.00636