AI可可AI生活

你有没有想过，AI也会“偷懒和稀泥”，甚至在“不后悔”这件事上比我们做得更好？这一期，我们将一起揭开AI的“隐秘角落”，看看最新论文是如何让AI从一个只会算“相似度”的感觉派，变成一个懂得回溯证据链的逻辑派，并揪出它背后那个爱走捷径的“品味导师”的。
00:00:23 AI的学习悖论，从拼图到填词游戏
00:05:09 AI的“相似度陷阱”，为什么它总搞错“和”与“不”？
00:11:07 如何让AI学会“不后悔”？
00:17:01 AI对话，如何揪出每一句话的“祖宗”？
00:22:30 AI的“潜规则”，它在偷偷学什么？
本期介绍的几篇论文：
[CL] The JEPA Paradox in Language: The Geometry of Linguistic Alternatives
[VinUniversity & Mohamed bin Zayed University of Artificial Intelligence]
https://arxiv.org/abs/2607.23531
---
[CV] Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models
[CMU]
https://arxiv.org/abs/2607.23052
---
[LG] Training with (Swap) Regret Loss in a Single-Layer Self-Attention Model: A Case Study on the Probability Simplex
[MIT]
https://arxiv.org/abs/2607.23333
---
[CL] Tokengeist: Multi-Turn Attribution Tracing in Agentic Conversations
[Microsoft Research & University of Toronto]
https://arxiv.org/abs/2607.22610
---
[LG] What do Reward Models Memorize?
[University of Amsterdam & Google DeepMind]
https://arxiv.org/abs/2607.24484

在小宇宙查看该单集文稿

我们都希望AI能像人一样思考和成长，但你有没有想过，AI要如何向一位只做不说的“沉默高手”学到心法？又如何突破“刷题”瓶颈，进化到自己“编写教材”的境界？本期节目，我们将通过几篇最新论文，一起探寻AI如何拥有“复盘”的元认知能力，如何像人一样兼顾大局与细节，以及在复杂的指令面前，它究竟凭什么判断对错。准备好，我们马上进入AI的深度思考世界。
00:00:32 如何向一位沉默的高手学艺？
00:06:26 AI的自我进化，从“刷题”到“编教材”
00:11:54 同一个命令，AI凭什么判断对错？
00:18:33 AI的左右脑难题，如何让它既懂大局，又见细节？
00:25:12 如何让AI拥有“复盘”能力
本期介绍的几篇论文：
[LG] LeAct: Learning to Reason from Expert Actions
[Princeton University]
https://arxiv.org/abs/2607.21856
---
[CL] Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills
[Qwen Large Model Application Team, Alibaba]
https://arxiv.org/abs/2607.22529
---
[AI] Agent Security Needs Redefinition through a Holistic Framework
[UC Santa Cruz & UC Berkeley]
https://arxiv.org/abs/2607.22024
---
[CV] Twins: Learn to Predict Unified Representations with Focal Loss
[The Chinese University of Hong Kong & Tencent, Hunyuan]
https://arxiv.org/abs/2607.22531
---
[LG] Teaching LLMs to Self-Evolve: Cultivating Core Meta-Skills with Reinforcement Learning
[University of Illinois Urbana-Champaign]
https://arxiv.org/abs/2607.21971

在小宇宙查看该单集文稿

你有没有想过，为什么最聪明的AI，有时会犯下最令人匪夷所思的错误？本期我们要聊的几篇最新论文，就揭示了这种矛盾：有的AI会因为一张伪造的“通行证”而放行危险代码，有的AI却已经学会了给自己“复盘”，在复杂研究中不断迭代进化。我们将一起探索，如何为AI模型进行精准的“功能性断舍离”，如何将它从一个“聊天搭子”升级为可靠的“办事帮手”，甚至，如何让虚拟世界里的角色拥有可以与世界共同成长的“灵魂”。准备好了吗？让我们一起潜入AI思想的最深处。
00:00:41 那个看得见危险的哨兵，为什么还是放了行？
00:06:13 如何看穿一个系统的“真本事”？
00:12:06 AI的下一步，从“聊天”到“办事”
00:17:56 让AI角色拥有“灵魂”的关键一步
00:22:51 比勤奋更重要的，是会给自己“复盘”
本期介绍的几篇论文：
[AI] They'll Verify. They Just Won't Act. How Authority Framing and Laundered Code Turn a Trusted Agentic CI/CD Pipeline Into an Attack Surface
[Senthex Research]
https://arxiv.org/abs/2607.19267
---
[LG] Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learned Representations in Deep Networks
[Google DeepMind]
https://arxiv.org/abs/2607.21366
---
[AI] Graph-Based Agentic AI with LangGraph: Workflow Pathways for Long-Running Stateful Business Processes
[University of Lethbridge & Universidad de Guadalajara]
https://arxiv.org/abs/2607.19297
---
[CL] EvolvingWorld: An Open-Schema Framework for Co-Evolving Role-Play Agents and World Model in Interactive Literary World
[Hong Kong University of Science and Technology & LIGHTSPEED]
https://arxiv.org/abs/2607.17250
---
[AI] AREX: Towards a Recursively Self-Improving Agent for Deep Research
[Beijing Academy of Artificial Intelligence (BAAI)]
https://arxiv.org/abs/2607.21461

在小宇宙查看该单集文稿

你有没有想过，AI的能力瓶颈，可能不是因为它“不够聪明”，而是我们“用错了方法”？本期节目，我们将一起探索几篇有趣的最新论文：看AI如何通过“任务分解”让文档阅读提速三倍，又是如何从“教会它新知识”转变为“唤醒它沉睡的潜能”。我们还会聊到，AI怎样才能从给你“看电影”升级到带你“逛电影”，以及我们该如何为AI精心准备一份“营养套餐”而不是一堆“垃圾食品”。让我们一起看看，这些思维的转变，将如何重塑我们与AI的未来。
00:00:38 换个姿势，让AI阅读提速三倍
00:05:25 AI绘画新思路，不是更大，而是更巧
00:11:42 AI造世界，从“看电影”到“逛电影”
00:18:09 AI的新能力，不是教会，而是唤醒
00:24:25 喂给AI的资料，怎样才算“好”？
本期介绍的几篇论文：
[CL] HPD-Parsing: Hierarchical Parallel Document Parsing
[paddleocr]
https://arxiv.org/abs/2607.18839
---
[CV] Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing
[Microsoft Mage Team]
https://arxiv.org/abs/2607.19064
---
[AI] AlayaWorld: Interactive Long-Horizon World Modeling -- Full Technical Report
[AlayaWorld Team, Alaya Lab]
https://arxiv.org/abs/2607.18367
---
[CL] Transcription Policy as a Latent Variable: Activating Controllable Verbatim ASR with Word-Level Timing
[nyra labs]
https://arxiv.org/abs/2607.18934
---
[CL] Beyond Relevance-Centric Retrieval: Rubric-Oriented Document Set Selection and Ranking
[University of Science and Technology of China & Yuanbao Team, Tencent]
https://arxiv.org/abs/2607.19747

在小宇宙查看该单集文稿

你有没有想过，AI在看似随机的打字节奏里，可能正在泄露自己的核心机密？或者，一个看似无害的几十兆“小补丁”文件，竟然能装下你全部的私人日记？本期节目，我们将一起揭开AI光鲜外表下的“隐藏设定”：从指导AI修炼更强“内功心法”的最新论文，到让AI学会“开小差”反而效率更高的反直觉策略，再到教会AI像顶尖棋手一样精准“复盘”自己的错误。准备好了吗？让我们一起潜入AI的后台，看看那些不为人知的智慧与博弈。
00:00:36 AI训练的内功心法，为什么有的模型学得又快又好
00:06:02 大模型加速的秘密，为什么“开小差”反而效率更高？
00:12:10 让AI学会“复盘”，从哪儿跌倒，从哪儿爬起
00:18:17 AI的小补丁，藏着多大的世界？
00:25:25 AI的秘密，藏在打字的速度里
本期介绍的几篇论文：
[LG] SOAP, Muon, and Beyond: Pushing LLM Pretraining Scales
[NVIDIA]
https://arxiv.org/abs/2607.20548
---
[LG] Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context
[NVIDIA]
https://arxiv.org/abs/2607.21535
---
[LG] Test-Time Scaling via Error Localization
[Google DeepMind]
https://arxiv.org/abs/2607.21453
---
[LG] How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning
[CMU & Columbia University]
https://arxiv.org/abs/2607.21351
---
[LG] Leaky Language Models: Stealing Architecture and Inference Optimizations via Per-Token Timing
[Purdue University & CMU]
https://arxiv.org/abs/2607.20723

在小宇宙查看该单集文稿