你有没有想过,AI要如何像高手一样,同时“试驾”多种思路?我们又该如何给狂飙的AI装上“定速巡航”,让它在学习时永不“翻车”?今天,我们就从几篇最新的AI论文出发,聊一聊AI要如何学会“分身术”思考,如何跳出“思维定式”的陷阱,甚至,我们以后可能再也不用费劲地给AI设定KPI,直接“说人话”就能让它们完美协作。准备好了吗?让我们一起探索AI思考方式的深层变革。 00:00:35 如何像高手一样思考?答案可能在“分身术”里 00:05:07 给狂飙的AI装上定速巡航 00:09:57 思维定式是怎么炼成的?AI给了我们一个新答案 00:15:23 怎么让AI大模型学会“左右互搏”? 00:21:37 AI界的“KPI”革命,未来我们不用再跟机器打哑谜 本期介绍的几篇论文: [CL] Multiplex Thinking: Reasoning via Token-wise Branch-and-Merge [Microsoft Research & University of Pennsylvania] https://arxiv.org/abs/2601.08808 --- [LG] Controlled LLM Training on Spectral Sphere [Microsoft Research Asia & Renmin University] https://arxiv.org/abs/2601.08393 --- [LG] Rewarding the Rare: Uniqueness-Aware RL for Creative Problem Solving in LLMs [MIT & NUS] https://arxiv.org/abs/2601.08763 --- [LG] Reverse Flow Matching: A Unified Framework for Online Reinforcement Learning with Diffusion and Flow Policies [MIT] https://arxiv.org/abs/2601.08136 --- [LG] The End of Reward Engineering: How LLMs Are Redefining Multi-Agent Coordination [New York University & Lerna AI] https://arxiv.org/abs/2601.08237
你有没有想过,一个更聪明的AI,是应该更会“思考”,还是更会“偷懒”?最新论文告诉我们,让AI学会用“记忆”分担计算,反而能让它更专注于难题。当AI面对一本几十万字的小说时,它又是如何像我们一样“做笔记”,避免“七秒记忆”的?更有趣的是,如果把AI关进小黑屋,不给任何学习资料,它竟能通过“左右互搏”实现自我进化。最后,我们会深入AI的内心世界,看看它“一本正经胡说八道”时,脑子里究竟走了哪两条路,以及它那令人惊叹的“举一反三”,可能根本不是在学习,而是在“对答案”。 00:00:48 为什么“偷懒”的AI,反而更会思考? 00:06:15 AI的“七秒记忆”,有救了? 00:11:46 AI的“闭关修炼”,不喂数据,如何变强? 00:16:53 AI为什么会“一本正经地胡说八道”?它的脑子里有两条路 00:22:20 别再说AI在“学习”了,它可能只是在“对答案” 本期介绍的几篇论文: [LG] Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models [DeepSeek-AI] https://arxiv.org/abs/2505.11080 --- [LG] Gecko: An Efficient Neural Architecture Inherently Processing Sequences with Arbitrary Lengths [University of Southern California & Meta AI Research] https://arxiv.org/abs/2601.06463 --- [LG] Dr. Zero: Self-Evolving Search Agents without Training Data [Meta Superintelligence Labs] https://arxiv.org/abs/2601.07055 --- [CL] Two Pathways to Truthfulness: On the Intrinsic Encoding of LLM Hallucinations [Peking University & Microsoft Research Asia] https://arxiv.org/abs/2601.07422 --- [LG] Filtering Beats Fine Tuning: A Bayesian Kalman View of In Context Learning in LLMs [UC Berkeley] https://arxiv.org/abs/2601.06100
本期我们来聊聊AI世界里那些“反直觉”的智慧:当AI不再给商品打分而是直接“写”出排名,当语音助手不再被粗暴地对答案而是被“手把手”教会思考,当“不完美”的数据反而能帮我们做出更好的决策,一场关于效率和认知的革命正在悄然发生。最新论文告诉我们,解决难题最好的方法,有时是换一个全新的玩法。 00:00:28 你看到的结果,是谁为你排的序? 00:07:42 AI大模型背后,一场关于“搬家”的效率革命 00:13:08 你的语音助手,为什么一开口就变笨了? 00:18:01 AI的“思考开关”,是个美丽的误会? 00:23:25 别再等了!“不完美”的数据也能做出好决策 本期介绍的几篇论文: [IR] Autoregressive Ranking: Bridging the Gap Between Dual and Cross Encoders [Google DeepMind & University of Massachusetts Amherst] https://arxiv.org/abs/2601.05588 --- [LG] MoEBlaze: Breaking the Memory Wall for Efficient MoE Training on Modern GPUs [Meta Platforms Inc] https://arxiv.org/abs/2601.05296 --- [CL] Closing the Modality Reasoning Gap for Speech Large Language Models [Microsoft Corporation & The Chinese University of Hong Kong] https://arxiv.org/abs/2601.05543 --- [LG] Do Sparse Autoencoders Identify Reasoning Features in Language Models? [UC Berkeley] https://arxiv.org/abs/2601.05679 --- [LG] Good Allocations from Bad Estimates [Stanford University & Max Planck Institute for Intelligent Systems, Tübingen] https://arxiv.org/abs/2601.05597
你有没有想过,我们能否打造一个既有“文科生”的灵活,又有“理科生”严谨的AI?当一群“偏科”的AI专家聚在一起,如何才能组建一支高效的“梦之队”?本期节目,我们将一口气为你解读几篇最新论文,看看科学家们是如何通过巧妙的流程设计,让AI学会“左右脑”分工、进行词级别的精细协作,甚至拥有主动管理记忆的“断舍离”能力。最后,我们还会揭秘一份顶尖科学家的“创新食谱”。准备好了吗?让我们一起探索AI进化背后的智慧。 00:00:38 AI的“左右脑”,如何让它既灵活又靠谱 00:06:27 AI也“偏科”?我们如何组建一个“梦之队” 00:11:25 AI当科学家,为什么还是个“学徒”? 00:20:09 让AI学会“断舍离”,它才能真正进化 00:25:12 科学家的创新,原来是有“食谱”的 本期介绍的几篇论文: [LG] Structured Decomposition for LLM Reasoning: Cross-Domain Validation and Semantic Web Integration [Warsaw University of Technology] https://arxiv.org/abs/2601.01609 --- [CL] Token-Level LLM Collaboration via FusionRoute [Meta AI] https://arxiv.org/abs/2601.05106 --- [LG] Why LLMs Aren't Scientists Yet: Lessons from Four Autonomous Research Attempts [Lossfunk] https://arxiv.org/abs/2601.03315 --- [CL] Agentic Memory: Learning Unified Long-Term and Short-Term Memory Management for Large Language Model Agents [Alibaba Group] https://arxiv.org/abs/2601.01885 --- [LG] Sci-Reasoning: A Dataset Decoding AI Innovation Patterns [Orchestra Research] https://arxiv.org/abs/2601.04577
你有没有想过,你每天使用的AI,可能正悄悄地把一整本《哈利波特》藏在“脑子”里?为了让AI变得更强,我们竟然要逼它和它所有的前辈“打群架”?本期节目,我们将一起揭开AI那些不为人知的“秘密”:从一个能让AI拥有完美记忆的“文件柜”,到一个既聪明又省钱的“免疫系统”,再到一场揪出AI“作弊考生”的全新考试。准备好了吗?让我们一起窥探AI大脑的奇妙内部。 00:00:32 你的AI,可能藏着一个图书馆 00:05:07 为什么说“盯住第一”是最大的陷阱? 00:11:19 给AI安一个靠谱的“文件柜” 00:16:50 给AI装上一个“既聪明又省钱”的免疫系统 00:23:26 你的AI考了高分,但它真的看懂图了吗? 本期介绍的几篇论文: [CL] Extracting books from production language models [Stanford University] https://arxiv.org/abs/2601.02671 --- [LG] Digital Red Queen: Adversarial Program Evolution in Core War with LLMs [MIT] https://arxiv.org/abs/2601.03335 --- [LG] Everything is Context: Agentic File System Abstraction for Context Engineering [University of New South Wales & ArcBlock, Inc & University of Tasmania] https://arxiv.org/abs/2512.05470 --- [LG] Constitutional Classifiers++: Efficient Production-Grade Defenses against Universal Jailbreaks [Anthropic] https://arxiv.org/abs/2601.04603 --- [LG] DatBench: Discriminative, Faithful, and Efficient VLM Evaluations [DatologyAI] https://arxiv.org/abs/2601.02316
我们总希望AI更像一个聪明的伙伴,而不是一个笨拙的机器。但怎样才算“聪明”?本期节目,我们将透过几篇最新的研究,一起窥探AI学习智慧的深层秘密。我们会聊到,AI如何像婴儿一样,在无声的世界里自己“悟”出万物的规律;又如何像个特工,在“聊天模式”和“任务模式”间无缝切换;我们还会探讨,如何用一把精妙的尺子,量出AI学到的究竟是“真本事”还是“假把式”,以及如何避免它在多重目标下“偏科”,甚至沦为一个只会讨好规则的“马屁精”。 00:00:39 AI学会了“无师自通”,世界将有什么不同? 00:06:21 给AI装上一个“万能遥控器” 00:12:57 AI上课也分“顿悟”和“补课”?一把尺子量出它学到了多少真本事 00:19:54 AI“偏科”怎么办?谈谈多目标奖励的艺术 00:25:33 “好学生”与“马屁精”,AI如何学会做个人 本期介绍的几篇论文: [LG] Learning Latent Action World Models In The Wild [FAIR at Meta] https://arxiv.org/abs/2601.05230 --- [LG] XGrammar 2: Dynamic and Efficient Structured Generation Engine for Agentic LLMs [Shanghai Jiao Tong University & CMU] https://arxiv.org/abs/2601.04426 --- [LG] Excess Description Length of Learning Generalizable Predictors [UC Berkeley & Anthropic] https://arxiv.org/abs/2601.04728 --- [CL] GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization [NVIDIA] https://arxiv.org/abs/2601.05242 --- [CL] Learning to Simulate Human Dialogue [Stanford University] https://arxiv.org/abs/2601.04436
今天,我们将一同探寻,AI如何通过理解“点的流动”来获得物理直觉,又为何会掉进人类语言的“隐喻陷阱”。我们还会深入AI的大脑,看看它的知识是如何生长又被遗忘的,并学习一种给AI做“大脑针灸”的调教术,治好它的“固执病”。最后,我们将揭秘一项最新研究,看看AI是如何被教会用“道德”这把尺子,去读懂网络上的“阴阳怪气”的。 00:00:34 机器人如何获得“物理直觉”? 00:05:05 你以为AI很理性?其实它活在比喻里 00:10:54 AI的大脑里,知识是怎么“长”出来又“丢”掉的? 00:16:36 AI调教术,给模型做一次“大脑针灸” 00:22:42 教AI读懂“阴阳怪气”,靠的是什么? 本期介绍的几篇论文: [RO] PointWorld: Scaling 3D World Models for In-The-Wild Robotic Manipulation [Stanford University & NVIDIA] https://arxiv.org/abs/2601.03782 --- [CL] Metaphors are a Source of Cross-Domain Misalignment of Large Reasoning Models [The University of New South Wales & CSIRO Data61] https://arxiv.org/abs/2601.03388 --- [CL] How Do Large Language Models Learn Concepts During Continual Pre-Training? [UC Davis & Virginia Tech & UCLA] https://arxiv.org/abs/2601.03570 --- [CL] ContextFocus: Activation Steering for Contextual Faithfulness in Large Language Models [Adobe Research, India] https://arxiv.org/abs/2601.04131 --- [CL] Self-Explaining Hate Speech Detection with Moral Rationales [University of São Paulo & University of Southern California & Saarland University] https://arxiv.org/abs/2601.03481
今天我们要聊一个特别有意思的话题:AI训练是不是一定要“大力出奇迹”?本期节目将通过几篇最新的论文,带你探索AI如何从海量数据中提炼出真正的“结构性知识”,如何像老师傅一样“边干边学”适应新环境,甚至如何靠一个“黄金样本”就打通任督二脉。我们将一起见证,AI正从一个埋头苦干的“莽夫”,进化成一个懂得“刻意练习”的聪明手艺人。 00:00:30 数据里,能“算”出新东西吗? 00:07:38 让机器像人一样,边干边学 00:12:36 AI训练的省钱秘笈,一块显卡如何干出两块的活? 00:19:27 AI进化的秘密,从“大力出奇迹”到“一招鲜吃遍天” 00:24:11 AI的“手艺”是怎么练成的? 本期介绍的几篇论文: [LG] From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence [CMU & New York University] https://arxiv.org/abs/2601.03220 --- [LG] In-Context Reinforcement Learning through Bayesian Fusion of Context and Value Prior [University of Cambridge & Mila - Quebec AI Institute] https://arxiv.org/abs/2601.03015 --- [LG] Chronicals: A High-Performance Framework for LLM Fine-Tuning with 3.51x Speedup over Unsloth [N/A] https://arxiv.org/abs/2601.02609 --- [LG] One Sample to Rule Them All: Extreme Data Efficiency in RL Scaling [GAIR & Taobao & Tmall Group of Alibaba] https://arxiv.org/abs/2601.03111 --- [LG] From Memorization to Creativity: LLM as a Designer of Novel Neural-Architectures [University of Würzburg] https://arxiv.org/abs/2601.02997
你有没有想过,当AI不再是只会模仿的“鹦鹉”,它会如何为自己打造一张世界的“活地图”,甚至为万物创造出能自主思考的“数字分身”?最新论文揭示,AI正通过一系列奇妙的方法解决自己的“健忘症”与“数据饥荒”,甚至开始反思“堆料越多越糊涂”的怪圈。今天,我们就来聊聊AI是如何学会拥有“活地图”、创造“数字分身”、进行“模拟推理”,并最终实现自我“瘦身”的。 00:00:34 你的脑子里,是不是也有一张“活地图”? 00:05:55 你我皆有“数字分身”,当AI为万物造“镜像” 00:12:23 你的常识可能被颠覆了,模仿来的思考,算不算思考? 00:17:46 预测的难题,当AI遇上“数据饥荒” 00:24:46 AI大模型内卷,为什么堆料越多,脑子越糊涂? 本期介绍的几篇论文: [LG] Flow Equivariant World Models: Memory for Partially Observed Dynamic Environments [Harvard University & CMU] https://arxiv.org/abs/2601.01075 --- [AI] Digital Twin AI: Opportunities and Challenges from Large Language Models to World Models [Lehigh University & University of Maryland & University of New South Wales] https://arxiv.org/abs/2601.01321 --- [CL] Simulated Reasoning is Reasoning [RWTH Aachen University & CMU] https://arxiv.org/abs/2601.02043 --- [LG] Zero-shot Forecasting by Simulation Alone [Amazon] https://arxiv.org/abs/2601.00970 --- [LG] Geometric and Dynamic Scaling in Deep Transformers [New York University & Stony Brook University] https://arxiv.org/abs/2601.01014
本期节目,我们将深入AI的“引擎盖”之下,看看那些看不见的结构如何决定一切。你会听到,为何区区2%的数据就能决定翻译能力的生死;AI如何像侦探一样,为复杂问题画出“破案地图”;以及在看似无害的模型拼接中,如何暗藏着难以察觉的“木马”后门。准备好了吗?让我们一起探索这些最新论文背后,令人拍案叫绝的智慧。 00:00:31 AI翻译的秘密,2%的数据,50%的能力 00:05:35 你以为的“搜索”,正在被重新发明 00:13:00 为什么你的“笨办法”,却是AI的“开窍”法? 00:18:16 AI大厨做菜重复?换一种“盐”试试 00:23:43 AI世界的“乐高”游戏,藏着一个你没想到的“后门” 本期介绍的几篇论文: [CL] The Role of Mixed-Language Documents for Multilingual Large Language Model Pretraining [University College London & Nanyang Technological University & University of Waterloo] https://arxiv.org/abs/2601.00364 --- [CL] Retrieval--Reasoning Processes for Multi-hop Question Answering: A Four-Axis Design Framework and Empirical Trends [University of Pittsburgh & Google Cloud AI Research] https://arxiv.org/abs/2601.00536 --- [LG] Deep Networks Learn Deep Hierarchical Models [Hebrew University of Jerusalem] https://arxiv.org/abs/2601.00455 --- [CV] It's Never Too Late: Noise Optimization for Collapse Recovery in Trained Diffusion Models [UC Berkeley & University of Tübingen] https://arxiv.org/abs/2601.00090 --- [LG] The Trojan in the Vocabulary: Stealthy Sabotage of LLM Composition [Purdue University & CMU] https://arxiv.org/abs/2601.00065
你有没有想过,我们能不能让AI像探险家一样,在脑中绘制一张动态的世界地图?或者,仅仅是换个“看”图的顺序,就能让AI的识别能力大幅提升?本期节目,我们将一起探索几篇有趣的最新论文:看看为什么用“假奖励”瞎指挥,反而能激发AI的潜能;AI又是如何自动发现数据背后的“主线任务”;以及最关键的,我们如何教会AI那句宝贵的“我不确定”,让它变得更值得信赖。 00:00:35 你的大脑,如何给世界画地图? 00:05:33 AI识图的秘密,你以为不重要的,恰恰是关键 00:12:01 为什么瞎指挥也能练出好学生? 00:17:57 有一种AI,能自动发现数据的“主线任务” 00:23:29 给AI装上一个“靠谱”探测器 本期介绍的几篇论文: [LG] MapFormer: Self-Supervised Learning of Cognitive Maps with Input-Dependent Positional Embeddings [Institut Jean Nicod & École Normale Supérieur] https://arxiv.org/abs/2511.19279 --- [LG] REOrdering Patches Improves Vision Models [University of Pittsburgh & UC Berkeley] https://arxiv.org/abs/2505.23751 --- [LG] Spurious Rewards: Rethinking Training Signals in RLVR [University of Washington] https://arxiv.org/abs/2506.10947 --- [LG] Distributional Autoencoders Know the Score [University of Michigan] https://arxiv.org/abs/2502.11583 --- [LG] Similarity-Distance-Magnitude Activations [Reexpress AI] https://arxiv.org/abs/2509.12760
你有没有想过,真正的智能不只在于堆砌知识,更在于懂得“断舍离”,甚至学会如何“聪明地努力”?这一期,我们将看到最新论文如何教会AI进行动态的自我修正,以及机器人如何通过构建“知识金字塔”学会心灵手巧。我们还会见证,古老的“三角形”如何在AI新魔法的加持下重返巅峰,以及我们如何通过更换一把“度量尺”,让小模型的训练经验直接指导大模型。准备好,一场关于AI学习智慧的认知升级,马上开始! 00:00:38 AI学会了“断舍离”,才能变得更聪明 00:05:48 机器人学会“心灵手巧”的秘密,不止是苦练 00:10:47 三角形,凭什么重返巅峰? 00:16:26 AI进化的新姿势,从“大力出奇迹”到“聪明地努力” 00:22:33 训练AI,我们是不是一直在“蒙眼下山”? 本期介绍的几篇论文: [LG] Deep Delta Learning [Princeton University & University of California, Los Angeles] https://github.com/yifanzhang-pro/deep-delta-learning/blob/master/Deep_Delta_Learning.pdf --- [RO] GR-Dexter Technical Report [ByteDance] https://arxiv.org/abs/2512.24210 --- [CV] Triangle Splatting for Real-Time Radiance Field Rendering [University of Liège] https://arxiv.org/abs/2505.19175 --- [CL] ShinkaEvolve: Towards Open-Ended And Sample-Efficient Program Evolution [Sakana AI] https://arxiv.org/abs/2509.19349 --- [LG] Training Deep Learning Models with Norm-Constrained LMOs [EPFL] https://arxiv.org/abs/2502.07529
与播客爱好者一起交流
添加微信好友,获取更多播客资讯
播放列表还是空的
去找些喜欢的节目添加进来吧