1038 episodes
- 你有没有想过,让AI变得更聪明,关键可能不是让它知道得更多,而是教会它如何更高效地“思考”?本期我们要聊的几篇最新论文,就深入到了AI的思维深处:从让AI懂得“选择性遗忘”以实现长时间推理,到揭开决定AI学习成败的三个神秘“开关”。我们甚至会看到,机器人是如何通过“自言自语”来规划复杂任务的。准备好一起探索AI大脑的内部运作机制了吗?我们马上开始!
00:00:33 如何让AI长时间思考,还不“累”?
00:05:05 给你一个确定性的菜谱,靠谱吗?
00:10:44 你关心的问题,AI能比专家更快找到答案吗?
00:16:24 拆开AI的“黑箱”,决定它聪明的三个开关
00:22:50 机器人会思考,需要分几步?
本期介绍的几篇论文:
[CL] Prefix Sliding for efficient test-time scaling
[Stanford University & University of California at Santa Barbara & University of Washington]
https://arxiv.org/abs/2608.26070
---
[LG] Beyond Optimal Rates in Stochastic Optimization: Trajectory-Adaptive Stopping Rules
[UC Berkeley & PSL Research University]
https://arxiv.org/abs/2608.25551
---
[AI] Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings
[Google Research]
https://arxiv.org/abs/2608.26088
---
[LG] Demystifying Reinforcement Learning Post-Training of Language Models
[University of Washington]
https://arxiv.org/abs/2608.24949
---
[RO] R^3: Training Robots to Reason in Natural Language via Reinforcement Learning
[Carnegie Mellon University (CMU)]
https://arxiv.org/abs/2608.26053
在小宇宙查看该单集文稿 - 本期我们要聊点脑洞大开的:如果让一群AI自己组建科研社区,甚至给它们“放假”,会涌现出怎样的科学发现?我们会看到,AI真正的成长秘诀,不在于修正答案,而在于递归式地优化自己的“思考方法”,甚至学会像孙悟空一样用“分身术”同时探索多种可能。接着,当AI团队犯错时,我们将化身侦探,精准定位“责任人”,并揭秘一个让AI提速的妙招——不是靠堆算力,而是靠精明的“预算”分配。准备好了吗?让我们一起从几篇最新论文中,探寻这些关于AI工作流、团队协作与自我进化的深刻洞见。
00:00:43 AI也需要“放假”?科学发现的新模式
00:06:03 成长的秘密,不是优化答案,而是优化方法
00:10:58 让AI学会“分身术”,我们能快多少?
00:16:06 AI犯错,我们应该怪谁?
00:20:55 AI 为什么那么慢?这篇论文给了个巧妙的答案
本期介绍的几篇论文:
[AI] Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment
[DualverseAI & University of California San Diego]
https://arxiv.org/abs/2608.23691
---
[AI] Metan^n: Recursive Self-Improvement through Emergent Depth
[University of Minnesota & Seoul National University]
https://arxiv.org/abs/2608.24735
---
[AI] Parason: Revealing Subtask and Trial Parallelism in LLM Reasoning
[Tsinghua University & NVIDIA]
https://arxiv.org/abs/2608.24658
---
[CL] Who is the Agent to Blame? Localizing Faithfulness and Citation Mistakes in Agentic Deep Research
[Bar-Ilan University & UNC Chapel Hill]
https://arxiv.org/abs/2608.24306
---
[CL] AgentSpec: Speculative Decoding for Batch Inference of LLM Agents
[The Ohio State University & Microsoft Research & University of Michigan]
https://arxiv.org/abs/2608.24004
在小宇宙查看该单集文稿 - 你有没有觉得,最聪明的AI有时也会犯一些“低级错误”?本期节目,我们就从几篇最新论文出发,去看看AI那些意想不到的“脆弱时刻”。我们将一起探索,为什么AI合作有时会“1+1<2”,甚至被少数派“带偏”;又为什么一个不起眼的错别字,就能让它瞬间“走神儿”。更重要的是,我们将看到科学家们如何像一位“自动马鞍匠”一样,为AI打造不断进化的外部装备,又如何通过“字斟句酌”的反馈,教会AI抵御外界的恶意指令。
00:00:36 如何给AI装上一个“自动升级”的马鞍?
00:05:13 为什么笼统的批评没用?从教AI“防骗”的底层逻辑说起
00:10:36 1+1 < 2?合作的隐形成本
00:15:48 一个好汉三个帮,AI为何越帮越忙?
00:20:58 为什么一个错别字,就能让AI“走神儿”?
本期介绍的几篇论文:
[AI] AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces
[POSTECH & KAIST & Southern University of Science and Technology]
https://arxiv.org/abs/2608.23041
---
[AI] SecOPD: Mitigating Adaptive Prompt Injections by On-Policy Distillation
[UC Berkeley]
https://arxiv.org/abs/2608.21500
---
[CL] The Collaboration Tax: How Much LLM Multi-Agent Systems Pay to Coordinate
[University of Notre Dame & Meta Superintelligence Labs]
https://arxiv.org/abs/2608.22152
---
[CL] Aligned Alone, Misaligned Together: Forecasting Adversarial Capture in LLM Agent Populations
[ETH Zurich & Tel Aviv University]
https://arxiv.org/abs/2608.22444
---
[CL] Lexical Perturbations Disrupt LLM Reasoning: An Empirical Study of Attention Diversion
[Missouri University of Science and Technology & University of North Texas]
https://arxiv.org/abs/2608.22140
在小宇宙查看该单集文稿 - 你是否想过,如何给总爱“胡说八道”的AI配个从不说谎的“裁判”,让它成为绝对可靠的实干家?又该怎么让两个AI跳过聊天,直接“开个小会”高效同步想法?本期节目,我们将一起探索几篇最新论文带来的奇妙思路:看AI如何一边深度思考,一边给自己“抢答”来提升速度;看机器人如何通过专属“陪练”从失败中自我开窍;最后,我们还会发现,通往最优解的道路,有时竟是一条返璞归真的捷径。
00:00:34 让AI从“夸夸其谈者”变成“实干家”的秘密
00:05:54 当AI学会了开小会
00:10:23 让AI一边思考,一边抢答
00:14:53 给机器人请个“陪练”,让它自己开窍
00:19:03 优化世界的返璞归真之道
本期介绍的几篇论文:
[AI] AI with Authority, from Application to Silicon
[J Hickey]
https://arxiv.org/abs/2608.21356
---
[LG] Dual-Cache Latent Space Communication between Heterogeneous Language Models
[J Liu, Q Zhang, Y Jia, Z Kan… — Amazon Web Services (AWS)]
https://arxiv.org/abs/2608.20617
---
[CL] Self-Speculation for Faster Reasoning Models
[R Valluri, T Nguyen, A Grover — University of California, Los Angeles]
https://arxiv.org/abs/2608.20359
---
[RO] Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning
[V Giridhar, A Khandelwal, J A. Collins, I Georgiev… — Georgia Institute of Technology]
https://arxiv.org/abs/2608.21204
---
[LG] Primal Acceleration of Newton's Method
[N Doikov — Cornell University]
https://arxiv.org/abs/2608.21359
在小宇宙查看该单集文稿 - 你是否想过,给AI的“私教”偷看标准答案,究竟是作弊还是神操作?为什么机器人模仿完美师傅反而会碰壁,甚至有时需要“闭着眼睛”走路?本期节目,我们将从几篇最新论文出发,揭示AI如何从模仿走向探索,以及强大的模型是如何在你的手机里实现性能飞跃的。让我们一起探寻这些AI“反常识”行为背后的智慧吧!
00:00:29 AI界的“陪练”与“私教”
00:06:33 机器人学艺,师傅领进门,修行靠自己
00:11:55 机器人为什么要“闭着眼睛”走路?
00:18:01 你的手机,为什么能越来越“聪明”?
00:22:35 高手与笨蛋的分界线,在于如何面对复杂
本期介绍的几篇论文:
[LG] Le Critique: Privileged Value Functions for LLM Reinforcement Learning
[Mistral AI]
https://arxiv.org/abs/2608.16739
---
[RO] FetchMan: Learning Visual Humanoid Loco-Manipulation Policies from Simulated Experiences
[University of California, Los Angeles & Allen Institute for AI]
https://arxiv.org/abs/2608.17027
---
[RO] Revisiting Open-Loop Execution in Robotics: Toward Reactive, Higher-Performing Policies
[MIT & UC Berkeley]
https://arxiv.org/abs/2608.15938
---
[LG] FlashAttention for Scalable Vector Architectures
[Chalmers University of Technology & University of Glasgow]
https://arxiv.org/abs/2608.18656
---
[CV] Falcon Perception-HD: High Density Perception via Reinforcement Learning
[Technology Innovation Institute]
https://arxiv.org/abs/2608.18881
在小宇宙查看该单集文稿
More Technology podcasts
Trending Technology podcasts
About AI可可AI生活
来自 @爱可可-爱生活 的第一手AI快报,用最简单易懂的语言,带你直击最前沿的人工智能科研动态。无论你是科技小白,还是行业达人,这里都有你想知道的AI故事和未来趋势。跟着我们,轻松解锁人工智能的无限可能!
#人工智能 #科技前沿
Podcast websiteListen to AI可可AI生活, Acquired and many other podcasts from around the world with the radio.net app

Get the free radio.net app
- Stations and podcasts to bookmark
- Stream via Wi-Fi or Bluetooth
- Supports Carplay & Android Auto
- Many other app features
Get the free radio.net app
- Stations and podcasts to bookmark
- Stream via Wi-Fi or Bluetooth
- Supports Carplay & Android Auto
- Many other app features


AI可可AI生活
Scan code,
download the app,
start listening.
download the app,
start listening.





























