AI可可AI生活
Episodes
AI前沿：从梯度下降模拟提示到数据效能的革命
28 Jun 2025
Contributed by Lukas
[CL] Can Gradient Descent Simulate Prompting?[MIT CSAIL]https://arxiv.org/abs/2506.20989---[CL] Potemkin Understanding in Large Language Models[MIT &a...
[感悟]两把尺子，量出两种人生
27 Jun 2025
Contributed by Lukas
决定你人生品质的，不是世界怎么看你，而是你如何看待你自己的世界。
给大象绣花，换个针法就简单了？
27 Jun 2025
Contributed by Lukas
[LG] Orthogonal Finetuning Made Scalable[Max Planck Institute for Intelligent Systems & University of Cambridge]arxiv.org
AI公司的“裁员”秘密：谁是真正的骨干？
27 Jun 2025
Contributed by Lukas
[LG] Who Does What in Deep Learning? Multidimensional Game-Theoretic Attribution of Function of Neural Units[University Medical Center Eppendorf &...
AI换脑术：如何让模型一键学会新本事？
27 Jun 2025
Contributed by Lukas
[LG] Command-V: Pasting LLM Behaviors via Activation Profiles[CMU]https://arxiv.org/abs/2506.19140
AI的秘密：打碎了还能认出来？
27 Jun 2025
Contributed by Lukas
[CL] Broken Tokens? Your Language Model can Secretly Handle Non-Canonical Tokenizations[University of Washington]https://arxiv.org/abs/2506.19004
高手过招：接力赛为何胜过群英会？
27 Jun 2025
Contributed by Lukas
[LG] Chain-of-Experts: Unlocking the Communication Power of Mixture-of-Experts Models[Northwestern University]https://arxiv.org/abs/2506.18945
AI前沿：从代码生成到自动化科研
27 Jun 2025
Contributed by Lukas
[CL] DiffuCoder：Understanding and Improving Masked Diffusion Models for Code Generation[Apple]https://arxiv.org/abs/2506.20639---[LG] Language Model...
[感悟]你的“监狱”有多大，世界就有多大
26 Jun 2025
Contributed by Lukas
这个世界上最大的“监狱”，不是用钢筋水泥建造的，而是我们自己思维的产物。
AI界的“好学生”：不仅会答题，还会写参考文献
26 Jun 2025
Contributed by Lukas
[CL] Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models[Duke University & Meta]https://arxiv.org/abs/2506.17585
聪明人的“笨功夫”，和“笨人”的聪明劲
26 Jun 2025
Contributed by Lukas
[LG] In-Context Learning Strategies Emerge Rationally[Stanford University & Harvard University]https://arxiv.org/abs/2506.17859
AI变聪明的秘密：请个好管家，而不是多招人
26 Jun 2025
Contributed by Lukas
[LG] Routing Mamba: Scaling State Space Models with Mixture-of-Experts Projection[Microsoft]https://arxiv.org/abs/2506.18145
AI变聪明的秘密：刷新“脑回路”，而不是扩容“硬盘”
26 Jun 2025
Contributed by Lukas
[LG] The 4th Dimension for Scaling Model Size[University of Illinois at Urbana-Champaign & University of Toronto]https://arxiv.org/abs/2506.18233
AI大模型：大力真能出奇迹吗？
26 Jun 2025
Contributed by Lukas
[LG] These are Not All the Features You are Looking For: A Fundamental Bottleneck In Supervised Pretraining[Facebook AI Research (FAIR) at Meta & ...
[感悟]你的“倒霉”，是世界给你的“私教课”
25 Jun 2025
Contributed by Lukas
当你给出“新答案”的瞬间，你会发现，那道反复折磨你的题，就这么烟消云散了。
如何让AI“明辨是非”，而不是“投机取巧”？
25 Jun 2025
Contributed by Lukas
[LG] Robust Reward Modeling via Causal Rubrics[Google DeepMind]https://arxiv.org/abs/2506.16507
我们怎么知道，AI是真的“懂了”？
25 Jun 2025
Contributed by Lukas
[LG] Latent Concept Disentanglement in Transformer-based Language Models[Purdue University & University of Southern California]https://arxiv.org/a...
人多，真的力量大吗？
25 Jun 2025
Contributed by Lukas
[CL] When Does Divide and Conquer Work for Long Context LLM? A Noise Decomposition Framework[University of Chicago & Together AI]https://arxiv.org...
给AI做“CT扫描”：一份来自科学家的操作手册
25 Jun 2025
Contributed by Lukas
[LG] On the Theoretical Understanding of Identifiable Sparse Autoencoders and Beyond[Peking University & MIT]https://arxiv.org/abs/2506.15963
养一个“AI神童”，原来是个精细活儿
25 Jun 2025
Contributed by Lukas
[CL] EvoLM: In Search of Lost Language Model Training Dynamics[Harvard & Stanford & EPFL]https://arxiv.org/abs/2506.16029
[感悟]放下“自证”：成年人最高级的自由
24 Jun 2025
Contributed by Lukas
真正的自由，是从不再解释开始。当你不再需要向世界证明什么时，你才真正开始拥有了自己的人生。
AI画画，怎样才能“又快又好”？
24 Jun 2025
Contributed by Lukas
[CV] Align Your Flow: Scaling Continuous-Time Flow Map Distillation [NVIDIA] https://arxiv.org/abs/2506.14603
AI 的“读心术”：我们如何才能信任一个“聪明的大脑”？
24 Jun 2025
Contributed by Lukas
[LG] Taming Polysemanticity in LLMs: Provable Feature Recovery via Sparse Autoencoders [Yale University & Shanghai Jiao Tong University] &nb...
你的AI在偷偷“修炼”：通往无穷的平坦大道
24 Jun 2025
Contributed by Lukas
[LG] Flat Channels to Infinity in Neural Loss Landscapes [EPFL & Flatiron Institute] https://arxiv.org/abs/2506.14951
从死记硬背到融会贯通，AI的“开窍”秘籍
24 Jun 2025
Contributed by Lukas
[LG] GrokAlign: Geometric Characterisation and Acceleration of Grokking [Rice University & Brown University] https://arxiv.org/abs/250...
[感悟]高手过招：如何砍掉内心的“全都要”？
23 Jun 2025
Contributed by Lukas
真正的自由，从来不是拥有无限的选择权。
AI育儿经：学霸是“刷题”刷出来的，还是“试错”试出来的？
23 Jun 2025
Contributed by Lukas
[CL] AceReason-Nemotron 1.1: Advancing Math and Code Reasoning through SFT and RL Synergy [NVIDIA] https://arxiv.org/abs/2506.13284
AI黑箱里的一张新地图
23 Jun 2025
Contributed by Lukas
[LG] Random Matrix Theory for Deep Learning: Beyond Eigenvalues of Linear Models [Huazhong University of Science and Technolog & UC Berkeley...
AI程序员“封神”？别急，先看看“奥赛冠军”的体检报告
23 Jun 2025
Contributed by Lukas
[LG] LiveCodeBench Pro: How Do Olympiad Medalists Judge LLMs in Competitive Programming? [New York University & Princeton University] ...
顶尖高手过招，为何有时“笨办法”反而更有效？
23 Jun 2025
Contributed by Lukas
[LG] Is your batch size the problem? Revisiting the Adam-SGD gap in language modeling [Max Planck Institute for Intelligent Systems] https...
[感悟]如何干掉那个叫“明天再说”的自己？
22 Jun 2025
Contributed by Lukas
你会发现，决定你五年、十年后成为一个什么样的人，根本不是你那些写在纸上、存在手机里的宏伟计划，而是你...
AI黑箱？不如跟它聊聊天
22 Jun 2025
Contributed by Lukas
[LG] Because we have LLMs, we Can and Should Pursue Agentic Interpretability[Google DeepMind]https://arxiv.org/abs/2506.12152
用嘴造工具：AI定制进入“白话时代”
22 Jun 2025
Contributed by Lukas
[LG] Text-to-LoRA: Instant Transformer Adaption[Sakana AI]https://arxiv.org/abs/2506.06105
AI绘画的秘密：创造力竟然源于“学不会”？
22 Jun 2025
Contributed by Lukas
[LG] On the Closed-Form of Flow Matching: Generalization Does Not Arise from Target Stochasticity[CNRS]https://arxiv.org/abs/2506.03719
AI的记忆革命：如何只记重点，忘掉其它？
22 Jun 2025
Contributed by Lukas
[CL] Don't Pay Attention[Avey AI]https://arxiv.org/abs/2506.11305
[感悟]别再“收藏”了！把笔记变成一台思想印钞机！
21 Jun 2025
Contributed by Lukas
真正的知识管理，不是收藏，而是创造。
让AI自己给自己“立规矩”，结果会怎样？
21 Jun 2025
Contributed by Lukas
[LG] AutoRule: Reasoning Chain-of-thought Extracted Rule-based Rewards Improve Preference Learning[CMU]https://arxiv.org/abs/2506.15651
你的“土办法”过时了，AI正在打造“策略工具箱”
21 Jun 2025
Contributed by Lukas
[LG] HeurAgenix: Leveraging LLMs for Solving Complex Combinatorial Optimization Challenges[Microsoft Research Asia]https://arxiv.org/abs/2506.15196
AI大脑里的“万金油”神经元，是臭虫还是宝贝？
21 Jun 2025
Contributed by Lukas
[LG] Dense SAE Latents Are Features, Not Bugs[MIT & ETH Zürich]https://arxiv.org/abs/2506.156
AI界的“调参玄学”：一个被遗忘的旋钮
21 Jun 2025
Contributed by Lukas
[LG] Optimal Embedding Learning Rate in LLMs: The Effect of Vocabulary Size[UC Berkeley & Microsoft Research]https://arxiv.org/abs/2506.15025
AI的“读心术”：模型的大脑会出卖它的秘密吗？
21 Jun 2025
Contributed by Lukas
[CL] Approximating Language Model Training Data from Weights[Cornell University]https://arxiv.org/abs/2506.155
[感悟]如何成为一个高明的“刻意选择者”？
20 Jun 2025
Contributed by Lukas
不被外界的噪音裹挟，不被内心的惯性驱动，清清楚楚地知道自己要什么，不要什么，然后全力以赴，成为一个清...
AI的“强迫症”：为什么你一句话没说完，它就“疯了”？
19 Jun 2025
Contributed by Lukas
[CL] Sampling from Your Language Model One Byte at a Time[University of Washington]https://arxiv.org/abs/2506.14123
为什么“抓重点”让AI学得更快？
19 Jun 2025
Contributed by Lukas
[LG] Transformers Learn Faster with Semantic Focus[IBM Research]https://arxiv.org/abs/2506.14095
AI的“识字”革命：我们离“读懂”世界又近了一步？
19 Jun 2025
Contributed by Lukas
[CL] From Bytes to Ideas: Language Modeling with Autoregressive U-Nets[FAIR at Meta]https://arxiv.org/abs/2506.14761
AI的“选择困难症”：通往更高智慧的秘密通道
19 Jun 2025
Contributed by Lukas
[CL] Reasoning with Exploration: An Entropy Perspective[RUC & MSRA & SJTU]https://arxiv.org/abs/2506.14758
AI训练的“七分饱”智慧
19 Jun 2025
Contributed by Lukas
[LG] Less is More: Undertraining Experts Improves Model Upcycling[Université de Montréal & Concordia University]https://arxiv.org/abs/2506.14126
[感悟]你的效率，藏在截止日期里
19 Jun 2025
Contributed by Lukas
工作量会像气体一样膨胀填满给定的时间容器，因此主动设定紧迫的截止日期能激发最高效率，让你成为时间的主...
为什么你总觉得，没人能真正懂你？
18 Jun 2025
Contributed by Lukas
[LG] Wanting to Be Understood Explains the Meta-Problem of Consciousness[Google DeepMind]https://arxiv.org/abs/2506.12086
AI的“分身术”：为什么你的万能助手不会再轻易“忘事”了？
18 Jun 2025
Contributed by Lukas
[CL] Multipole Attention for Efficient Long Context Reasoning[UC Berkeley]https://arxiv.org/abs/2506.13059
AI的自我修炼：如何让机器拥有“反思”的能力？
18 Jun 2025
Contributed by Lukas
[CL] Direct Reasoning Optimization：LLMs Can Reward And Refine Their Own Reasoning for Open-Ended Tasks[Microsoft]https://arxiv.org/abs/2506.13351
我们用来分辨AI谎言的工具，它自己靠谱吗？
18 Jun 2025
Contributed by Lukas
[LG] Verifying the Verifiers：Unveiling Pitfalls and Potentials in Fact Verifiers[Yonsei University & Stanford University & University of Was...
给AI请个好家教：少即是多的学习智慧
18 Jun 2025
Contributed by Lukas
[CL] Refract ICL：Rethinking Example Selection in the Era of Million-Token Models[Google DeepMind]https://arxiv.org/abs/2506.12346
[感悟]你的人生，到底有多少“可能性”？
18 Jun 2025
Contributed by Lukas
人生的智慧不在于追求一步到位的完美答案，而在于始终为自己保留更多的选择权和可能性。
管中窥豹，竟能知全豹？
17 Jun 2025
Contributed by Lukas
[LG] Spectral Estimation with Free Decompression[UC Berkeley & University of Melbourne]https://arxiv.org/abs/2506.11994
AI 进化论：如何让机器像高手一样思考？
17 Jun 2025
Contributed by Lukas
[LG] TreeRL: LLM Reinforcement Learning with On-Policy Tree Search[Tsinghua University & California Institute of Technology]https://arxiv.org/abs/...
AI突然“开窍”？别激动，这可能不是你想的智能
17 Jun 2025
Contributed by Lukas
[CL] Large Language Models and Emergence: A Complex Systems Perspective[Santa Fe Institute]https://arxiv.org/abs/2506.11135
让AI长颗“人心”：机器如何学会人类的“偏见”？
17 Jun 2025
Contributed by Lukas
[LG] Tversky Neural Networks: Psychologically Plausible Deep Learning with Differentiable Tversky Similarity[Stanford University]https://arxiv.org/abs...
AI的“一招鲜”：如何只训练一次，就让模型十项全能？
17 Jun 2025
Contributed by Lukas
[CL] You Only Fine-tune Once：Many-Shot In-Context Fine-Tuning for Large Language Model[Google & University of Florida]https://arxiv.org/abs/2506...
[感悟]你活在谁的参照系里？
17 Jun 2025
Contributed by Lukas
现代人的焦虑很大程度源于不断的比较。我们习惯用别人的成就来衡量自己，这种"比较"就像快乐的小偷，偷走了本...
AI偷懒的艺术：如何用“草稿”撬动“大模型”？
17 Jun 2025
Contributed by Lukas
[CL] Draft-based Approximate Inference for LLMs[FuriosaAI & UW-Madison]https://arxiv.org/abs/2506.08373
AI世界的“社交难题”：光会拉黑还不够
17 Jun 2025
Contributed by Lukas
[LG] On the Similarities of Embeddings in Contrastive Learning[Yonsei University]https://arxiv.org/abs/2506.09781
AI的“鱼与熊掌”：既要跑得快，又要学得好？
17 Jun 2025
Contributed by Lukas
[LG] Sequential-Parallel Duality in Prefix Scannable Models[MIT CSAIL & Technical University of Munich]https://arxiv.org/abs/2506.10918
AI的终极进化：当机器学会给自己“划重点”
17 Jun 2025
Contributed by Lukas
[LG] Self-Adapting Language Models[MIT]https://arxiv.org/abs/2506.10943
AI的“开窍”秘诀：不是更聪明，而是会“偷懒”
17 Jun 2025
Contributed by Lukas
[LG] CoRT: Code-integrated Reasoning within Thinking[University of Science and Technology of China & Qwen Team & The Chinese University of Hon...
[论文品读]意图条件流占用模型
16 Jun 2025
Contributed by Lukas
[LG] Intention-Conditioned Flow Occupancy Models C Zheng, S Park, S Levine, B Eysenbach [Princeton University & UC Berkeley] 本...
[论文品读]用大语言模型求解不等式证明
16 Jun 2025
Contributed by Lukas
[LG] Solving Inequality Proofs with Large Language Models J Sheng, L Lyu, J Jin, T Xia... [Stanford University & UC Berkeley] 本...
[论文品读]强化学习教师的推理时扩展
16 Jun 2025
Contributed by Lukas
[LG] Reinforcement Learning Teachers of Test Time Scaling E Cetin, T Zhao, Y Tang [Sakana AI] 本文通过提出强化学习教师（...
[论文品读]分支薛定谔桥匹配
16 Jun 2025
Contributed by Lukas
[LG] Branched Schrödinger Bridge Matching S Tang, Y Zhang, A Tong, P Chatterjee [Duke-NUS Medical School & Quebec AI Institute]  ...
[论文品读]扩散二象性
15 Jun 2025
Contributed by Lukas
[LG] The Diffusion Duality S S Sahoo, J Deschenaux, A Gokaslan, G Wang, J Chiu, V Kuleshov [Cornell Tech & EPFL Lausanne] 本文...
AI前沿：从并行思维到超快规划
13 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的四项前沿研究，带来耳目一新的洞见：1.《Multiverse: Your Language Models Secretly Decid...
AI前沿：从推理外推到医疗影像
12 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究，带来耳目一新的洞见： “e3: Learning to Explore Enables Extrapolation o...
AI前沿：从神经网络学习到智能体行动
11 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究，涵盖了神经网络学习理论、语言模型训练、优化技术、模型效率...
AI前沿：从自适应模型到关系推理的未来
10 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究，揭示了模型自适应、效率提升及复杂数据处理的新突破。包括：...
AI前沿：从模型思维到代码进化
09 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五项AI前沿研究的关键进展：1.《When Does Closeness in Distribution Imply Representational Similarity? ...
AI前沿：AI推理的幻觉与突破
08 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究，揭示了AI推理能力的真实面貌与优化策略，并展望了机器人技术...
AI前沿：从模型嫁接到遗忘之谜
07 Jun 2025
Contributed by Lukas
本期“TAI快报”深入探讨了五篇AI前沿论文的关键内容：1.《Exploring Diffusion Transformer Designs via Grafting》提出了“嫁接...
AI前沿：从长文本生成到无人驾驶
06 Jun 2025
Contributed by Lukas
本期“TAI快报”深入探讨了AI领域的五项前沿研究：1. “Rectified Sparse Attention”通过周期性校准解决长文本生成中的...
AI前沿：从语言偏差到训练谜团
05 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了人工智能领域的五项前沿研究，揭示了AI模型设计与训练中的隐藏挑战与创新突破。首先...
AI前沿：从语言适配到智能体思维
04 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了人工智能领域的五大前沿研究，涵盖语言模型适配、能力评估、智能体思维、训练策略...
AI前沿：从分钟级游戏学习到语言模型的记忆之谜
03 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五项AI领域的最新研究成果，带来前沿技术洞见：1. “Test-Time Training Done Right” 通过大块...
AI前沿：从Adam到偏好学习的性能之谜
02 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究，涵盖优化算法、强化学习、学术工具、脑启发计算和偏好学习。...
AI前沿：从分布比较到语言模型的“游荡”本质
01 Jun 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI领域的前沿论文，揭示了多项关键进展： 《Kernel Quantile Embeddings and Associated Proba...
AI前沿：隐秘动力与高效学习
31 May 2025
Contributed by Lukas
本期“TAI快报”深入探讨了五篇AI领域的前沿论文，揭示了模型内部机制与优化策略的新视角。包括：通过动力系统...
AI前沿：从熵管理到长思维链的秘密
30 May 2025
Contributed by Lukas
本期“TAI快报”深入探讨了AI推理能力的五大前沿研究，揭示了提升AI“思考”能力的新路径。包括：通过熵管理解...
AI前沿：从语言操控到自我反思
29 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI领域的前沿论文，揭示了语言模型操控、推理反思、自我训练及多语言能力增强的...
AI前沿：小模型也能聪明，数据选择有大招
28 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究：1.《Small Models, Smarter Learning: The Power of Joint Task Training》揭示联合...
AI前沿：从数据污染检测到高效推理
27 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究：1.《How Can I Publish My LLM Benchmark Without Giving the True Answers Away?》提...
AI前沿：从几何对称到推理控制
26 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI领域的前沿论文，带来以下关键洞见： AdS-GNN - a Conformally Equivariant Graph Neural Netw...
AI前沿：从自适应思考到像素推理
25 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五项AI前沿研究：1.《Thinkless: LLM Learns When to Think》提出自适应推理框架，让语言模型根...
AI前沿：从“快思慢想”到文本水印的前沿探索
24 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五项AI前沿研究：1. “Incentivizing Dual Process Thinking for Efficient Large Language Model Reasoning”揭...
AI前沿：从软性推理到自设计智能体的突破
23 May 2025
Contributed by Lukas
本期“TAI快报”深入探讨了AI领域的五项前沿研究，涵盖文本生成、推理优化、用户反馈学习、训练课程设计和多智...
AI前沿：从慢思考到高效推理
22 May 2025
Contributed by Lukas
本期《TAI快报》聚焦AI“思考”艺术，深入探讨了五项前沿研究：1.《Reward Reasoning Model》提出AI在评价前先“思考”...
AI前沿：从破碎表征到高效计算的突破
21 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI领域的前沿论文，带来耳目一新的洞见。首先，“Questioning Representational Optimism in...
AI前沿：从语言对齐到游戏建模
20 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了AI领域的五项前沿研究，涵盖语言模型对齐、代码优化、图像生成和游戏世界建模等多个...
AI前沿：AI如何突破多语言、稀疏学习与药物设计的边界
19 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了四篇AI领域的前沿论文，涵盖多语言多模态、稀疏函数学习、分子设计和时间感知等方向...
AI前沿：AI如何颠覆数学、音乐与经济
18 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI领域的前沿论文，揭示了AI在数学、音频生成、经济分析、数据筛选及分布式训练...
AI前沿：从热力学到进化论
17 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI前沿论文的关键洞见： 《Neural Thermodynamic Laws for Large Language Model Training》提出神...
AI前沿：从自我奖励到因果推理的突破
16 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI前沿论文，涵盖语言模型的自主学习、神经网络在线学习、上下文处理机制、机器...
AI前沿：从困惑到推理解锁语言模型的秘密
15 May 2025
Contributed by Lukas
本期《TAI快报》深入探讨了五篇AI前沿论文，揭示了大语言模型的概率一致性、推理能力、效率优化与对齐机制的最...