![[人人能懂AI前沿] AI的瘦身术、组织图与紧箍咒](https://image.xyzcdn.net/FuDP4HpAp8ezgVZMmEel3mblKCmJ.jpg@middle)
主播
节目简介
来源：小宇宙
你有没有想过，我们如何给AI这头“吞金兽”来一次彻底的瘦身和压缩？如何为它设计一张分工明确的“组织架构图”，让“调度员”和“专家”各司其职？我们又该如何给它装上一个既能规避灾难性风险，又能动态调整预算的“安全大脑”？当AI自己当上“裁判”时，我们如何确保它不是在抛硬币？本期节目，我们将通过几篇最新的研究，一起探索如何让AI变得更高效、更聪明，也更可靠。
00:00:31 驯服AI的新兵法，“共享”与“压缩”
00:06:36 给AI画一张“组织架构图”，谁是调度员，谁是专家？
00:13:07 如何让AI既能干，又不出事？
00:18:11 AI当裁判，是明察秋毫，还是抛硬币？
00:23:49 给AI上好“紧箍咒”，它才能学得又快又稳
本期介绍的几篇论文：
[LG] Gefen: Optimized Stochastic Optimizer
[Reichman University & Tel Aviv University]
https://arxiv.org/abs/2606.13894
---
[LG] A theoretical model for task routing in mixture-of-expert transformers
[University of Sydney & Zhejiang University]
https://arxiv.org/abs/2606.14398
---
[LG] Utility-Constrained Policy Optimization
[York University & Google DeepMind]
https://arxiv.org/abs/2606.14029
---
[CL] The Coin Flip Judge? Reliability and Bias in LLM-as-a-Judge Evaluation
[A Yagubyan]
https://arxiv.org/abs/2606.13685
---
[LG] Diffusion Policy Optimization without Drifting Apart
[UC Berkeley]
https://arxiv.org/abs/2606.13795
00:00:31 驯服AI的新兵法，“共享”与“压缩”
00:06:36 给AI画一张“组织架构图”，谁是调度员，谁是专家？
00:13:07 如何让AI既能干，又不出事？
00:18:11 AI当裁判，是明察秋毫，还是抛硬币？
00:23:49 给AI上好“紧箍咒”，它才能学得又快又稳
本期介绍的几篇论文：
[LG] Gefen: Optimized Stochastic Optimizer
[Reichman University & Tel Aviv University]
https://arxiv.org/abs/2606.13894
---
[LG] A theoretical model for task routing in mixture-of-expert transformers
[University of Sydney & Zhejiang University]
https://arxiv.org/abs/2606.14398
---
[LG] Utility-Constrained Policy Optimization
[York University & Google DeepMind]
https://arxiv.org/abs/2606.14029
---
[CL] The Coin Flip Judge? Reliability and Bias in LLM-as-a-Judge Evaluation
[A Yagubyan]
https://arxiv.org/abs/2606.13685
---
[LG] Diffusion Policy Optimization without Drifting Apart
[UC Berkeley]
https://arxiv.org/abs/2606.13795
外观
