你有没有想过，AI也会有“选择困难症”吗？或者，怎么才能给AI请个既省钱又能干的“陪练”？这一期，我们就来聊聊几篇有趣的最新论文，看看科学家们是如何教会AI像高手一样反思、像侦探一样倾听“沉默的投票”，甚至用中学物理知识，给AI装上一双“3D眼睛”的。准备好了吗？让我们一起出发！

00:00:29 鸡娃不如“陪练”，AI训练的降本增效新思路

00:05:53 AI的学霸秘籍，如何像高手一样思考和进化

00:12:12 AI的阅读术，如何既快又好地啃下海量信息？

00:18:24 AI的“过分自信”，原来是种“选择困难症”

00:23:39 让AI拥有“立体视觉”的省钱妙计

本期介绍的几篇论文：

[LG] CoDistill-GRPO: A Co-Distillation Recipe for Efficient Group Relative Policy Optimization

[Google Research]

https://arxiv.org/abs/2605.08873

---

[CL] RubricEM: Meta-RL with Rubric-guided Policy Decomposition beyond Verifiable Rewards

[Google Cloud AI Research]

https://arxiv.org/abs/2605.10899

---

[CL] Scratchpad Patching: Decoupling Compute from Patch Size in Byte-Level Language Models

[Google DeepMind]

https://arxiv.org/abs/2605.09630

---

[CL] The Silent Vote: Improving Zero-Shot LLM Reliability by Aggregating Semantic Neighborhoods

[Google]

https://arxiv.org/abs/2605.09739

---

[LG] RelFlexformer: Efficient Attention 3D-Transformers for Integrable Relative Positional Encodings

[Seoul National University & Google Research]

https://arxiv.org/abs/2605.10706