你有没有想过，让AI变聪明，或许并不需要更强的算力，而是需要一种更巧妙的引导方式？本期，我们将一起探索几篇最新论文带来的奇妙洞见：我们会发现，一点点“计算噪声”竟能让AI学得更好；我们甚至能像做CT扫描一样，亲眼“看见”AI思考的几何轨迹；并学习如何像教育孩子一样，教会AI在探索与专注间找到完美平衡，甚至不花一分钱，就解锁它的隐藏潜能。00:00:36 不花钱升级你的AI？换个提问方式就行00:05:39 AI育儿经：如何教机器学会“恰到好处”的探索00:11:50 训练AI，加点“噪声”效果更好？00:16:47 AI的“心流”：看见思考的轨迹00:22:19 如何让聪明的AI，学会更聪明地做事？本期介绍的几篇论文：[LG] Reasoning with Sampling: Your Base Model is Smarter Than You Think[Harvard University]https://arxiv.org/abs/2510.14901---[LG] Agentic Entropy-Balanced Policy Optimization[Kuaishou Technology & Renmin University of China]https://arxiv.org/abs/2510.14545---[LG] QeRL: Beyond Efficiency -- Quantization-enhanced Reinforcement Learning for LLMs[NVIDIA & MIT]https://arxiv.org/abs/2510.11696---[LG] The Geometry of Reasoning: Flowing Logics in Representation Space[Duke University]https://arxiv.org/abs/2510.09782---[CL] Demystifying Reinforcement Learning in Agentic Reasoning[National University of Singapore & Princeton University & University of Illinois at Urbana-Champaign]https://arxiv.org/abs/2510.11701
--------
30:11
--------
30:11
[人人能懂] 从统一语言、认知陷阱到犯错的艺术
要让AI真正变聪明，是该为它发明一套统一江湖的“世界语”，还是该教会它如何在“大脑袋”和“长思考”间做出明智的权衡？又或者，我们得先帮它建立健康的“信息食谱”，避免陷入“脑子瓦特”的陷阱，并治愈它那藏着世界性偏见的好奇心？甚至，最新论文告诉我们，通往更高智能的钥匙，竟然是先让AI学会如何当一个会犯错的“差生”。本期节目，我们将通过五篇最新论文，共同探索AI智能背后那些你意想不到的深层逻辑。00:00:39 AI的“世界语”：一套统一江湖的武功秘籍00:06:11 AI的“垃圾食品”陷阱：为什么顶尖模型也会“脑子瓦特”？00:10:58 AI变聪明，靠脑子大还是想得久？00:17:18 AI的“好奇心”，藏着一个世界性的偏见00:22:56 为什么最聪明的AI，要先学会当个“差生”？本期介绍的几篇论文：[LG] Tensor Logic: The Language of AI [University of Washington] https://arxiv.org/abs/2510.12269 ---[CL] LLMs Can Get "Brain Rot"! [Texas A&M University & University of Texas at Austin & Purdue University] https://arxiv.org/abs/2510.13928 ---[LG] Not All Bits Are Equal: Scale-Dependent Memory Optimization Strategies for Reasoning Models [KRAFTON & University of Wisconsin–Madison] https://arxiv.org/abs/2510.10964 ---[CL] The Curious Case of Curiosity across Human Cultures and LLMs [University of Michigan] https://arxiv.org/abs/2510.12943 ---[LG] Learning to Make MISTAKEs: Modeling Incorrect Student Thinking And Key Errors [MIT CSAIL] https://arxiv.org/abs/2510.11502
--------
28:31
--------
28:31
[人人能懂] 从火箭发射、大学主修到片刻沉思
你有没有想过，除了喂给它更多数据，还有哪些更精妙的法门能让AI变得更聪明？本期我们要聊的几篇最新论文，就揭示了AI的“成长秘籍”：它们把训练AI的视角从“下山”升级为“发射火箭”，为它设计了从通识到专业的“大学课程”，还教会了它预测“未来摘要”的远见，以及在关键时刻“喘口气”慢思考的智慧。今天，就让我们一起看看，这些研究是如何重塑AI的“学习方法论”的。00:00:33 训练AI，你以为是爬山，其实是开火箭？00:05:56 AI成长秘籍：多上一门“专业课”00:11:26 AI模型的终极瘦身术：如何让大象既轻盈又聪明？00:16:53 AI的远见：不只关心下一个词00:21:10 AI的“沉思时刻”：快与慢的智慧本期介绍的几篇论文：[LG] Optimal Control Theoretic Neural Optimizer: From Backpropagation to Dynamic Programming[Meta & Georgia Institute of Technology & Apple]https://arxiv.org/abs/2510.14168---[CL] Midtraining Bridges Pretraining and Posttraining Distributions[CMU]https://arxiv.org/abs/2510.14865---[LG] BitNet Distillation[Microsoft Research]https://arxiv.org/abs/2510.13998---[LG] Beyond Multi-Token Prediction: Pretraining LLMs with Future Summaries[FAIR at Meta & CMU]https://arxiv.org/abs/2510.14751---[CL] Catch Your Breath: Adaptive Computation for Self-Paced Sequence Production[Google DeepMind]https://arxiv.org/abs/2510.13879
--------
26:08
--------
26:08
[人人能懂] 从科学预测、大道至简到团队协作
想知道为什么教机器人玩最“笨”的玩具，反而能让它学会抓取任何东西吗？本期节目，我们将一起探索如何将神秘的AI“炼金术”变成一门严谨的科学，看看怎样让AI大神学会“说人话”并带得动AI小白，并最终揭示，那些五花八门的调教秘籍背后，其实藏着同一个简单的目标。让我们马上进入今天的前沿速递！00:00:28 AI大模型调教指南：从玄学到科学00:05:39 返璞归真：最笨的方法，可能就是最好的方法00:11:25 想让机器人变聪明？先教它玩“笨”玩具00:16:41 如何让AI大神，带得动AI小白？00:00 大模型调教秘籍：条条大路通罗马？本期介绍的几篇论文：[LG] The Art of Scaling Reinforcement Learning Compute for LLMs[Meta & UT Austin & UC Berkeley]https://arxiv.org/abs/2510.13786---[RO] VLA-0: Building State-of-the-Art VLAs with Zero Modification[NVIDIA]https://arxiv.org/abs/2510.13054---[RO] Learning to Grasp Anything by Playing with Random Toys[UC Berkeley]https://arxiv.org/abs/2510.12866---[LG] Tandem Training for Language Models[Microsoft & EPFL & University of Toronto]https://arxiv.org/abs/2510.13551---[LG] What is the objective of reasoning with reinforcement learning?[University of Pennsylvania & UC Berkeley]https://arxiv.org/abs/2510.13651
New Zealand