[{"data":1,"prerenderedAt":5258},["ShallowReactive",2],{"post-\u002Fblog\u002F2026\u002F2026-08-13-xie-zuo-yu-yan-de-zhuan-bian":3,"i-boxicons:location":5250,"i-mingcute:time-line":5254,"i-ci:tag":5256},{"post":4,"nextPost":212,"prevPost":485},{"id":5,"title":6,"body":7,"description":13,"draft":198,"enableComment":199,"extension":200,"image":195,"important":198,"location":201,"meta":202,"navigation":199,"ogImage":203,"path":204,"seo":205,"stem":206,"summary":207,"tags":208,"time":210,"__hash__":211},"blog\u002Fblog\u002F2026\u002F2026-08-13-xie-zuo-yu-yan-de-zhuan-bian.md","写作语言的转变",{"type":8,"value":9,"toc":194},"minimark",[10,14,17,20,176,179,182,185,188,191],[11,12,13],"p",{},"长期以来，我对自己的博客定位从来不是分享，而更多是给自己看的「笔记」。很多情况是：记录一个刚刚搞懂的东西 → 写下来 → 以后自己查 → 写完就结束。我的博客基本都很短，基本三分钟以内就能读完。大约在 2023 年开始，同时为了学习英语，锻炼技术写作能力，我在疯狂背单词的同时，也可以保持用英语进行技术写作。现在坚持了快两年。所以长期以来，利用英语写作，对我来说压力并不大。写作时不需要追求完整论证，只需要把自己的思路编码下来。哪怕句子稍微生硬一点，只要自己以后能看懂，就已经达成目的。",[11,15,16],{},"过去两年的英语写作之所以轻松，是因为它服务于「笔记」功能——短、线性、结论明确。这种写作不需要复杂的逻辑嵌套，不需要反复回看和修改结构，英语的线性特征反而成了一种约束，帮你把想法压缩成清晰的短句。",[11,18,19],{},"但是后来，情况发生了变化，我的博客从「给自己看的笔记」的定位，开始转移到了「表达与推演」。当写作从记录变成表达与推演，语言就从工具变成了负担。写作负荷不是恒定的，它会随着结构复杂度和语言熟练度的乘积而非线性增长。",[11,21,22,23,175],{},"文章越来越长，关于技术写作的部分，有的文章我需要插入大量的 ",[24,25,29,52],"span",{"className":26,"translate":28},[27],"katex","no",[24,30,33],{"className":31},[32],"katex-mathml",[34,35,37],"math",{"xmlns":36},"http:\u002F\u002Fwww.w3.org\u002F1998\u002FMath\u002FMathML",[38,39,40,47],"semantics",{},[41,42,43],"mrow",{},[44,45,46],"mtext",{},"LaTeX",[48,49,51],"annotation",{"encoding":50},"application\u002Fx-tex","\\LaTeX",[24,53,57],{"className":54,"ariaHidden":56},[55],"katex-html","true",[24,58,61,66],{"className":59},[60],"base",[24,62],{"className":63,"style":65},[64],"strut","height:0.8988em;vertical-align:-0.2155em;",[24,67,71,76,81,113,117],{"className":68},[69,70],"mord","text",[24,72,75],{"className":73},[69,74],"textrm","L",[24,77],{"className":78,"style":80},[79],"mspace","margin-right:-0.36em;",[24,82,85],{"className":83},[84],"vlist-t",[24,86,89],{"className":87},[88],"vlist-r",[24,90,94],{"className":91,"style":93},[92],"vlist","height:0.6833em;",[24,95,97,102],{"style":96},"top:-2.905em;",[24,98],{"className":99,"style":101},[100],"pstrut","height:2.7em;",[24,103,105],{"className":104},[69],[24,106,112],{"className":107},[69,74,108,109,110,111],"mtight","sizing","reset-size6","size3","A",[24,114],{"className":115,"style":116},[79],"margin-right:-0.15em;",[24,118,120,124,128,167,171],{"className":119},[69,70],[24,121,123],{"className":122},[69,74],"T",[24,125],{"className":126,"style":127},[79],"margin-right:-0.1667em;",[24,129,132,158],{"className":130},[84,131],"vlist-t2",[24,133,135,153],{"className":134},[88],[24,136,139],{"className":137,"style":138},[92],"height:0.4678em;",[24,140,142,146],{"style":141},"top:-2.7845em;",[24,143],{"className":144,"style":145},[100],"height:3em;",[24,147,149],{"className":148},[69],[24,150,152],{"className":151},[69,74],"E",[24,154,157],{"className":155},[156],"vlist-s","​",[24,159,161],{"className":160},[88],[24,162,165],{"className":163,"style":164},[92],"height:0.2155em;",[24,166],{},[24,168],{"className":169,"style":170},[79],"margin-right:-0.125em;",[24,172,174],{"className":173},[69,74],"X"," 公式、推导和演绎过程。此时再使用非母语进行技术写作时心智负担显著增大。短文里，一句话如果不知道怎么写，一般情况可以用简单句子绕一下，问题不大。但是到了长篇技术文章，可能连续几页都在描述一个复杂的思想。这时候需要同时控制数学符号、技术概念、论证结构、前后术语一致性、句法，还有段落之间的衔接。于是工作记忆里面同时跑着很多东西。",[11,177,178],{},"而且，长篇博客并不是一次写完的，很可能需要花上几个星期慢慢打磨，或者经常回顾我自己的博客的思想。写作的时候，我大多数时间都在思考：「这个理论到底应该怎么解释？」。但是，结果脑子里却不断出现：「这里应该用 which 还是 that？」、「这个东西应该叫 representation 还是 formulation？」、「这个词性是否合适，有没有对应的名词形式？」之类细枝末节的语法问题。久而久之，写作起来会非常累。在短文本写作里，这种差异几乎感觉不到。但当文本从几百字增长到几千、几万字以后，语言的视觉结构、信息密度、词法形态、定位效率和工作记忆负担都会开始成为写作系统的一部分。",[11,180,181],{},"另外一种原因，英文是线形文字，阅读效率天然就很低：想在长文中定位到某一块位置，必须要从每一段从头开始逐行扫过，无法像汉字那样逐块扫描。很多时候我在长文写作中，需要不断地往回头看。而英语的语法线性强，从句嵌套多了以后，读者和作者都容易迷失。因为汉字同时包含视觉图像和声音两种信息，它信息密度和视觉辨识特征，使中文文本非常适合视觉扫描。而英文单词之间存在大量空格，真正有语义重量的东西被拆成了很多视觉单元。所以当文章达到几千甚至上万字以后，回来看自己几周前写的东西，会出现一种很奇怪的体验：中文是在「看结构」，英文更容易变成「读句子」。所以，就导致语义单元和视觉单元不对齐——一个概念可能要三四个单词才能表达，视觉上要扫过更长的距离才能抓住一个完整意群。",[11,183,184],{},"其实不止是我，很多技术写作者都会遇到一个问题：语言不是中性的容器，它会反过来塑造你思考的形状。",[11,186,187],{},"笔记型写作的特点，思维已经完成。只是在编码一个已经清晰的结论，供未来的自己检索。表达与推演型写作则不同，写作本身就是思维过程。你在写的过程中推演、发现、修正。文字不是思维的镜像，而是思维的工具。当写作成为思考工具时，语言就不再只是输出端的问题，而是输入端的问题。非常需要语言来帮助你组织尚未成形的想法，来试探逻辑的边界，来连接不同的概念。",[11,189,190],{},"从此以后，我的博客功能，从以前的博客 externalized memory（外部记忆），到现在逐渐变成了 externalized thinking（外部化思考）。",[11,192,193],{},"也许从英语写作到中文写作，是在为思维本身让路。让认知资源从语言操作中解放出来，还给真正的思考。",{"title":195,"searchDepth":196,"depth":196,"links":197},"",2,[],false,true,"md","河南郑州",{},null,"\u002Fblog\u002F2026\u002F2026-08-13-xie-zuo-yu-yan-de-zhuan-bian",{"title":6,"description":13},"blog\u002F2026\u002F2026-08-13-xie-zuo-yu-yan-de-zhuan-bian","长期以来，为了学习英语并增强熟练度，在技术写作时，我都会刻意使用英语来写作。但是后来我发现使用英语的阅读和思考心智负担非常大。从笔记到表达与推演，写作目标发生变化后，认知资源也需要实现对应的重新分配。",[209],"thoughts","2026-08-13T00:00:00+08:00","7ebbJ5U5Sd2pnIWInH_6aFeptiKE6s-ogM9nTwBT0EQ",{"id":213,"title":214,"body":215,"description":219,"draft":198,"enableComment":199,"extension":200,"image":195,"important":198,"location":201,"meta":477,"navigation":199,"ogImage":203,"path":478,"seo":479,"stem":480,"summary":481,"tags":482,"time":483,"__hash__":484},"blog\u002Fblog\u002F2026\u002F2026-08-14-demystification-research.md","祛魅科研，每个研究生的必修课",{"type":8,"value":216,"toc":469},[217,220,223,227,230,233,236,247,250,253,256,259,262,265,268,271,274,277,281,284,287,290,297,305,308,311,314,317,320,326,329,332,335,338,343,346,349,352,355,361,364,367,374,377,380,383,386,389,392,398,401,404,407,410,413,416,419,422,425,432,436,443,448,451,454,457,460,463,466],[11,218,219],{},"从 2025 年开始入学读研，到现在已经有一个年头了，不知不觉已经研二了。在过去的一年的研究生学习中，我发现我的心态、世界观经历了一次完整的、打击性的重塑。如果说之前我的心态是中规中矩，那么现在则是满心失落。自从经历过完整的科研训练，调研课题、找 idea，做实验、处理数据、撰写论文、修改文章、等投稿、写 Rebuttal、拒稿再转投、论文改版等一系列流程完整经历过后，只剩下满心苦涩。一度怀疑，我们搞的这个所谓的「科研」的价值和意义。😇",[11,221,222],{},"因为笔者读的研究生是计算机工程类专业，对于其他理工科领域、文科领域，笔者不是很熟悉，所以以下观点体验仅适用于该专业。",[224,225,226],"h2",{"id":226},"草台班子",[11,228,229],{},"第一个对我的改变就是对学术成果的怀疑。学术圈的科研，其实并没有那么高端，可以说是草台班子。",[11,231,232],{},"在读研之前，我一直认为学术论文其实是非常高端的东西，至少是可信度很高的参考。但是读研之后，这个想法被改变了：绝大多数论文，他们的观点、数据、实验，看看就好，不必认真。",[11,234,235],{},"一篇论文的结论，往往只有作者自己的机器上成立。审稿人不会复现，编辑不会复现，引用者也不会复现。大家都默认了这个不可验证的游戏潜规则。论文并不是经过验证的知识，而是一个经过美化的故事。",[11,237,238,239,246],{},"有人专门对被 ICML 收录的 Oral 论文进行了复现，并给出了",[240,241,245],"a",{"href":242,"rel":243},"https:\u002F\u002Fx.com\u002FChenhaoTan\u002Fstatus\u002F2079969545629118737",[244],"nofollow","论文的报告信息","。结果发现，在 105 篇完整的复现论文中，只有 27 篇复现了超过 40% 的声称效果，剩余的几乎无法复现，要么效果不达标。原因包括代码无法运行、文件缺失以及模型已弃用等等。",[11,248,249],{},"复现困难确实是机器学习领域长期存在的问题。因为，CUDA 版本、显卡型号、CPU 型号、PyTorch 框架版本，可能都有大量隐蔽的细节差异，如果再组合起来，那么效果就完全无法验证。实验中的很多关键 Hack（比如特定的随机种子、数据预处理的微妙顺序、超参数的精细调节）可能只存在于作者的实验笔记或潜意识里。这些隐性经验很难通过文字传递，导致复现者像在黑暗中摸索。",[11,251,252],{},"甚至如果作者的代码里有 Bug，也很可能会造成完全错误的结论。我曾经跑过前人的基线实验，发现一篇论文，他的某个指标又非常虚高。最后才发现，作者对数据集根本没有清洗干净，也就是说，这份成绩是假的。",[11,254,255],{},"但问题是，这个论文已经被大量后人引用且作为基线了，后人想要再投稿，那么他的成绩必须要比这个假成绩更高，但是你根本比不过这假成绩，怎么办呢？你只能偷偷摸摸用点魔法手段，万不可认真，否则反而会显得像是你做错了，审稿人不接受。后来的研究者反而成了受害者。这不就是难为后人么？",[11,257,258],{},"所以，后人也只能不得不逼着也参与这种数字游戏，偷偷使用手段。结果就是，学术论文逐渐沦为一场“数字通胀”。每一篇新论文都在前人的数字泡沫上再吹一层泡沫，直到某个数据集上的准确率达到 99.9%，然后这个方向就“死掉”了——因为没人能再超越了，而大家都知道那个 99.9% 是假的，但谁都不想捅破。",[11,260,261],{},"这已经超出了学术不端的范畴，演变成了一种系统性的「囚徒困境」，很遗憾，这种困境，在学术圈里几乎无解。",[11,263,264],{},"无法复现不一定是作者故意有造假意向，而是因为深度学习本身就是一个巨大的黑盒，可验证性、可解释性非常差。很多错误，作者不一定能及时发现，审稿人也不一定能发现，于是，“只要能 work 就能发文”。当审稿人面对一篇充满“惊艳”实验数据的论文时，他既没有算力去复现，也没有理论工具去证伪其内部的混沌过程。于是，审稿的标准退化为：“只要故事逻辑自洽，且结果看起来比 SOTA（当前最优）高，我就信。”这就给出了巨大的浑水摸鱼的空间。",[11,266,267],{},"这种无法复现的现象，也会给自己带来很多的麻烦：如果你的文章需要依赖前人的工作，但是前人的工作本身就有问题，那么你的工作就几乎不可能顺利下去。你只能花大把时间来反思、调试，最后还很可能一无所获。",[11,269,270],{},"最后被浪费大把的时间和精力。这种消耗是精神凌迟。后续工作往往不得不替它买单。",[11,272,273],{},"有时候，我也甚至怀疑作者是故意在开源里缺斤少两，删减关键代码和文件或者埋入 Bug，或者论文被 Accept 后立即下架数据集，极力避免复现。😇要知道，审稿人通常只有 2-3 位，他们在几周内免费审稿，不可能复现你的实验，也无法验证你的原始数据。他们主要检查的是逻辑是否顺畅和方法是否看起来合理，而不是结论是否绝对正确。",[11,275,276],{},"学术论文它的首要目标是发表，而不是传世。为了发表，作者必须讲一个完整且自洽的故事。他们会突出最漂亮的数据，而弱化不支持的“噪音”结果。所以，他们会在讨论部分夹带私货，把故事讲得更有吸引力。它也确实是痛苦的。因为这意味着你失去了一个可以无条件信任的知识权威。",[224,278,280],{"id":279},"ai-审稿危机","AI 审稿危机",[11,282,283],{},"读研起到现在，我已经投稿了三篇文章，这点我是有亲身体会的。",[11,285,286],{},"另一种尴尬是 AI 时代的审稿危机。AI 时代的恶果就是，论文灌水越来越容易。往常，写一篇论文很可能需要一学期、大半年的时间，要辛辛苦苦做实验、分析结果、画图表、斟酌写作。但是现在不一样了，利用自动化学术 Agent 如 AutoResearch，只要输入合适的提示词和模糊的想法再交给 Agent，它两天时间就可以编出一篇像模像样的论文，自动编写程序做实验，自动画图表，一个人一个月就能编写出 10 篇论文，直接投到顶会轰炸。",[11,288,289],{},"AAAI 在 2026 年收到了五万多份取号，几乎是 2025 年的两倍。而前年 2024 年才不过九千多份，短短两年时间就增长了五倍！其实不止 AAAI，其他的计算机会议的投稿量也是几乎翻倍增长。假设每篇论文标准配备 3 名审稿人，组委会将需要管理 120,000 份独立审稿意见：",[11,291,292],{},[293,294],"img",{"alt":295,"src":296},"计算机顶会近五年投稿量。自从 2025 年后，大多数投稿量都是翻倍式增长","https:\u002F\u002Fimage-assets.dreams.plus\u002F202608111219721.jpg",[11,298,299,300,304],{},"这就会导致一个后果：劣币驱逐良币。审稿体系正在加速崩溃。就算你辛辛苦苦完成了心血，设计实验，写好文章，",[301,302,303],"strong",{},"审稿人也几乎不会认真看你的文章，而是直接交给 LLM 敷衍。"," 因为审稿人压根无法区分哪篇论文是 AI 写得，哪篇不是。更糟糕的是，学术领域是高度分化的，审稿人极有可能也不熟悉、也不理解你的研究领域。所以，他也只能同样一股脑交给 AI 审稿。",[11,306,307],{},"在这种情况下，灌水零成本，中间大量灌水的作者和敷衍的审稿人占多数。认真的审稿人和认真的作者只能被动深受其害。受害的永远只是认真搞学术的你。在这种情况下，审稿工作还能正常进行下去吗？",[11,309,310],{},"而 LLM 审稿本身就有很大的问题，AI 并不会理解你的文章，LLM 的幻觉问题众所周知。它会用非常刁钻的角度在你的文章里挑刺，哪怕是没有问题也会制造出一些问题。这是因为 LLM 在训练、SFT 偏好微调的时候，天生就倾向于给你打低分。LLM 并不真正拥有论文作者的研究上下文。它尤其容易犯一种很危险的错误：把“我没理解”转换成“作者的方法存在问题”。",[11,312,313],{},"LLM 模型学习到的统计规律是：“审稿”这个动作的语义空间，几乎完全由“批评性词汇”构成。因此，即使一篇论文毫无瑕疵，模型根据概率分布生成的“审稿风格”文本，天然就带有负面倾向。SFT 微调的时候，宽松的论文意见并不受欢迎，为了要尽可能压榨 LLM 能力，在微调的时候会特意设计出非常严苛的训练案例。这种奖惩机制直接塑造了模型的“人格”：它必须“生产批评”来证明自己的价值。哪怕没有真实缺陷，它也会启动“防御性挑刺”模式，利用模式匹配强行构造出看似合理的问题。",[11,315,316],{},"我的文章被拒稿过一次，三个 Reviewer 里两人给出的意见，明显就是 AI 复制粘贴的。😅 但是你不能抗议，你还要捏着鼻子忍着恶心，在 rebuttal 里面假模假样地感谢审稿人，再假模假样地写 rebuttal，跪求他赏赐一口饭吃。😅",[11,318,319],{},"审稿体系存在一个非常明显的权力不对称。Reviewer 可以随意评价“The novelty is insufficient.”而作者很难说：“你根本没读懂我的论文。”因为作者没有证据。另外，Reviewer 可以洋洋洒洒写出几千字的审稿意见，没有字数限制，而作者给的 Rebuttal 却严格限制在几百字以内。这也是最恶心的一点😅",[11,321,322],{},[293,323],{"alt":324,"src":325},"如图，是几个 LLM 在审稿 ArXiv 计算机学科论文时给出的平均得分（10 分制）","https:\u002F\u002Fimage-assets.dreams.plus\u002F202608111210591.png",[11,327,328],{},"试过把 21-23 年，GPT 出来以前的三大顶会里，已经发表收录的论文随机爬下来，投给 AI 审，50 篇里 30 多个 weak reject，10 多个 weak accept，剩下的全部都是 reject。一篇 accept 都没有。后 AI 时代所谓的顶会论文已经变成笑话哩。",[11,330,331],{},"审稿人因为反驳文太长而疲惫，作者因为审稿人固执而绝望。当作者知道审稿人是 AI，审稿人知道作者用了 AI 时，作者 - 审稿人 - 编辑三方之间的这场学术对话，就彻底沦为了一场荒诞剧。这就造成了非常滑稽的景象，AI 写 AI 审：",[11,333,334],{},"用 AI 写论文、写代码，再用 AI 初审，根据 AI 的意见修改，完成初稿。审稿人拿到初稿，再交给 AI 审稿，用 AI 的意见给出 review 意见，作者拿到 AI 写的意见后再交给 AI 写 rebuttal。审稿人再用 AI 根据 rebuttal 做出决定。堪称学术出版领域正在逼近的赛博朋克式奇观。",[11,336,337],{},"在三方中，编辑\u002F主席的地位最为尴尬。当所有文字意见如审稿、反驳都由 AI 生成时，编辑失去了判断学术创新性的任何抓手。他唯一能做的，就是检查流程是否走完：AI 是否提了 3 个问题？作者是否逐条回复？回复长度是否达标？只要格式合规，就可以按下接受键。学术判断的权力，在此刻已完全让渡给了硅基算法。",[11,339,340],{},[301,341,342],{},"所以，能否发顶会、顶刊，实际上已经越来越像摸彩票中奖的运气、概率问题，跟你的文章质量、工作效果已经几乎没有什么关系了。",[224,344,345],{"id":345},"手艺人",[11,347,348],{},"硕博生这个身份曾经是不少人引以为傲的根本。但是真正体验过他们的生活，也会发现他们本质上和流水线的螺丝工人相差无几，这种生活状态其实是非常压抑的，如果过这种日子，那只能用「熬」来形容。",[11,350,351],{},"硕博生本质上也是出卖高强度脑力劳动换取生存资格的劳动者。跟工地上抗水泥的农民工、顶着大太阳湿透衣服的清洁工没有本质区别。",[11,353,354],{},"对于理工科研究生，996 是常态，实验室的灯永远亮着。每天睁眼闭眼就是要面对屏幕上密密麻麻的实验数据、代码和仪器。高强度脑力劳动后，带给身心的除了疲劳还是疲劳。日子久了，精气神会被消磨掉，慢慢丧失对这个世界的好奇心和一切欲望，不想谈恋爱，不想出去旅游，哪怕是手里的游戏，日子久了玩起来也没意思。",[356,357,358],"blockquote",{},[11,359,360],{},"到周末后，只想在床上躺着，啥也不干。脑一旦被单一的高强度任务长期占据，负责发散思维、感受情绪、产生欲望的脑区就会被持续抑制。只像一个漂浮在数据海洋里的意识，拖着一具沉重而麻木的肉体，犹如冢中枯骨而已。",[11,362,363],{},"这是最致命的。当你看清你所做的研究可能只是学术游戏里的一个废棋，对外部世界毫无影响时，熬就变成了一种精神上的凌迟。不禁会问，「我做的这一切，受了那么多折磨，到底有什么意义？」",[11,365,366],{},"工人进厂时，起码能自知之明、清醒地知道这是出卖身体，用劳动换生存。但硕博生被社会、被家人、被曾经的自己赋予了天子骄子、知识精英的光环。当现实变成日复一日地跑数据、伺候仪器时，日子久了你会有这样一种感觉：自己并不是一个人，而是一个庞大机器中的一颗零件。",[11,368,369,370,373],{},"硕士生（Master）也不是大师。博士生也不博学。他的知识广度，甚至还可能不如一个高中生。",[301,371,372],{},"读研后，你的视角只会被限制在高度狭窄且专业的小领域里。"," 在自己的学术孤岛里自娱自乐。",[11,375,376],{},"中世纪的经院哲学家热衷于争论「一个针尖上能站几个天使」，而今天的学术圈，大量精力被消耗在维护主流范式上。正如前文所提到的，学术论文本身也是高度固定化的、范式的。有时候你不会感觉自己「在创造一个想法」，而是「完成一个八股文」。",[11,378,379],{},"如果要比喻硕博生这个群体的身份，它更像是一个高度程序化的手艺人、螺丝工。搞科研的流程本身就是高度固定化、流水线化的。调研、idea、实验、写作、改稿、Rebuttal。这一圈下来，恭喜你，你已经是一名合格的劳工！",[224,381,382],{"id":382},"学术圈",[11,384,385],{},"曾经对学术圈的浪漫想象，至少代表了知识的前沿、先进的生产力和思想文化，具有进步性。我其实对科研领域内的学者，高校里的教授、教师等群体，长期是存在敬仰的，认为他们或多或少都是代表了人类开拓认知知识、征服星辰大海的一批人，在心里也会敬三分。",[11,387,388],{},"现在才知道，学术圈其实是高度封闭的。因为知识本身就有很强的入门壁垒，当人类认知突破到一定边界时，工具、术语和范式的复杂度必然形成门槛，学术方向往往会走向高度分化、隔行如隔山的细碎分支，各个分支又会高度壁垒。这就造成了学术圈的高度封闭性。",[11,390,391],{},"其实，越是高度封闭的圈子，越容易产生高度固化的权力结构，行事作风越是封建化、越是讲政治。😅其实现在的学术圈其实跟欧洲中世纪的经学院教派之争、西藏喇嘛们的辩经并没有什么区别。那些专家，领域学者，头衔看得是挺唬人，但做的无非在极度封闭的圈子里，用只有内部人能懂的黑话，争论着对外部世界影响甚微的问题，而决定胜负的常常不是真理，而是资历、人脉和对经典的诠释权。",[11,393,394,397],{},[301,395,396],{},"学术圈，其实比大多数人想象的还要小。"," 因为现代学科已经进入高度分化、高度专业化的时代了。如果细分下去，全中国十四亿人里，同一个领域的研究同行很可能不超过百人，甚至十几人。在一个村落里，任何小动作，全落在这几十个低头不见抬头见的人手里。这里没法对事不对人，因为在结构上，所有的事，最终都是人的事。",[11,399,400],{},"然而，现在的学术圈是零和博弈，资源是极其有限的。这个在申请基金、文章版面、学术交流等等活动中，只要有人胜出，必然会有人落选。这也就意味着，你的小圈子里，可能到处都在「树敌」。这并不意味着本人有错，而是你的存在，本身就是威胁。",[11,402,403],{},"所以遇到同行暗中使绊子，也是常见的事。你的审稿人很可能跟你的导师有竞争或者过节，就直接轻松 Reject 你的心血。即使现在的审稿制度大多是双盲制，在一个领域只有几十上百人的圈子里，根本不存在真正的双向匿名。看研究问题、看方法、看引用的文献，审稿人闭着眼都能猜到这篇稿子出自哪个课题组。",[11,405,406],{},"对一个埋头苦干的学生来说，这是最深的打击。你相信公正，相信学术质量至上。然后，一堵由学派、人情、资源争夺构成的墙，悄无声息地挡在你面前，将你的心血轻松驳回。你甚至没有一个明确的敌人去质问抗争，你甚至不知道你的敌人是谁，又得罪了谁，只剩下无尽的无力感和被戏弄的愤怒。这种被暗算的体验会深刻腐蚀对学术共同体的信任。",[11,408,409],{},"像中世纪的领主分封土地一样，大牛导师和顶尖实验室把持着顶级期刊的版面、重大项目的经费和学阀圈子的话语权。你想在他的领地上发文章，就得遵循他的范式、引他的文章、甚至拜他的码头。学术圈的游戏规则是，正确不等于接受。投稿像一场赌博，审稿人的口味、当期版面、甚至运气，都比你那篇精心打磨的论文权重更高。",[224,411,412],{"id":412},"事业意义",[11,414,415],{},"虽然小时候有「长大要当科学家」这种理想，但是，个人认为「学术」这条路并不适合像我这样的平民子弟。没有充裕的家底和财力作为后盾，吃学术这碗饭，也是一种高风险职业。",[11,417,418],{},"我觉得那些在学术圈里搞研究的人，其实也挺可悲的。自己把大量的青春，时间，精力投入到自己的课题里，勉强能讨得经费，靠这个饭碗。因为，选择某个研究领域，在初期往往带有偶然性。但一旦投入，就成了无法回头的豪赌。赌的是这个方向在几十年内不被证伪、不被超越、不被认为是死胡同、不会没落。他用的是几十年的青春和精力做赌注，用最严谨的头脑，从事着一项本质上充满不确定性的高风险事业。",[11,420,421],{},"如果有人突然跳出来用新理论、新范式挑战他，或者推翻了他的观点课题，这无异于把他降维打击成了一块废品，弃之如敝履，这不是嘲讽，而是真实的悲剧。在高度职业化的学术圈里，一个人的身份、地位、自尊，都深深扎根于他那一亩三分地的研究课题。",[11,423,424],{},"当他的理论被推翻，在外人看来不过是一个观点被证伪。但对他而言，无异于整个学术人格被判处了死刑，在这个圈子里，会被迅速边缘化，从而判了死刑。他毕生构建的意义大厦，瞬间崩塌为一座废墟。这就是一种存在主义危机。",[11,426,427,428,431],{},"大多数普通人活下去本身就很难。因为 ",[301,429,430],{},"他们的人生，还有其他责任"," 。若压制住七情六欲，寒窗几十年，去碰学术圈，也未免太委屈了。如果一个人 25 岁读博士，30 岁左右博士毕业，然后经历博士后、非升即走、青年项目竞争，他可能在四十岁前都处于高度竞争状态。而同期进入工业界的人，可能已经积累了财富、住房和职业资本。",[224,433,435],{"id":434},"破局功利化读研","破局：功利化读研",[11,437,438,439,442],{},"在中国，虽然知识分子往往被冠以社会期待的光环，但是这也是一种负担和枷锁。请记住：我们是普通人尤其是出身平民家庭，我们首要目标是生存。在生存生计成为问题的时候，我们 ",[301,440,441],{},"没有义务背负太多期待","。",[356,444,445],{},[11,446,447],{},"沧浪之水清兮，可以濯吾缨；沧浪之水浊兮，可以濯吾足。",[11,449,450],{},"请卸下你的「学术羞耻感」。把研究生学历视为一份职业准入资格证，而非学术朝圣。学术职业是一种高风险选择，而不是所有人都必须承担的使命。",[11,452,453],{},"如何破局——功利化读研不失为一种出路。研究生必须思考：我读研的目的是为了什么？获得更好的就业门槛，暂时避开竞争激烈的就业市场，获得更多选择权。这完全是一种合理的人生规划。",[11,455,456],{},"首要的当然是毕业、混文凭，获得硕博的身份————这也是最现实、也最重要的一条。所以，在读研前期，你的一切目标是，必须以尽可能短的时间，完成最低毕业要求。大量水论文、蹭项目就足矣，不需要尽善尽美。科研嘛，也就那样，随便搞搞就行。",[11,458,459],{},"这是功利化读研真正的溢价所在。既然科研只求及格，那你必须把多出来的精力毫无愧疚地投入生存技能的构建。",[11,461,462],{},"在当前的环境下，「包装」比「做事」更重要，「数量」比「质量」更重要。不要把自己的全部人生价值绑定在学术成果上。科研如此，创业如此，艺术如此，很多长期主义事业都是如此。我不需要证明自己是英雄，我只需要把自己的人生过好。知识值得敬畏，但人的生命也值得敬畏。学术可以是人生的一部分，但不必成为人生的全部。",[11,464,465],{},"希望研究生们，不必神化科研、不必自我内耗、不必绑定学术理想，认清行业真相后，依然可以清醒活着、务实成长。",[11,467,468],{},"到最后，这种「虽千万人，吾往矣」，本身就有壮士断腕的秋风式悲凉，不是吗？😮‍💨",{"title":195,"searchDepth":196,"depth":196,"links":470},[471,472,473,474,475,476],{"id":226,"depth":196,"text":226},{"id":279,"depth":196,"text":280},{"id":345,"depth":196,"text":345},{"id":382,"depth":196,"text":382},{"id":412,"depth":196,"text":412},{"id":434,"depth":196,"text":435},{},"\u002Fblog\u002F2026\u002F2026-08-14-demystification-research",{"title":214,"description":219},"blog\u002F2026\u002F2026-08-14-demystification-research","自 2025 年入学至今，一年光阴悄然而逝，我已步入研二。回望这一年的研究生生活，我的心态与世界观经历了一场彻底而沉重的重塑——曾经的从容平实，如今已被挥之不去的失落感取代。",[209],"2026-08-14T00:00:00+08:00","utt_zVmdYpEoW-2HkgNeH6tsTcym6PstUAxslVH-wUE",{"id":486,"title":487,"body":488,"description":5239,"draft":198,"enableComment":199,"extension":200,"image":195,"important":198,"location":203,"meta":5240,"navigation":199,"ogImage":203,"path":5241,"seo":5242,"stem":5243,"summary":5244,"tags":5245,"time":5248,"__hash__":5249},"blog\u002Fblog\u002F2026\u002F2026-08-04-the-representations-in-llm.md","大语言模型中的表征流形",{"type":8,"value":489,"toc":5229},[490,499,502,511,514,517,520,523,836,1186,1289,1293,1296,1440,1736,1739,1742,2353,2991,2994,2997,3000,3017,3020,3023,3026,3029,3625,3628,3701,3704,3707,3710,4150,4355,4561,4822,4895,4898,4901,4904,4907,4947,5014,5097,5100,5103,5106,5109,5112,5115,5126,5129,5132,5135,5138,5141,5144,5147,5220,5223,5226],[11,491,492,493,498],{},"如果你关注了过去两年机制可解释性研究的进展，很可能见过这样一类图：把大语言模型某一层神经元的激活向量投影到二维或三维空间并作图，会发现「月份」「星期几」「年份」「颜色」等概念的表示并不是杂乱无章地散布在空间中，而是沿着一条优美的曲线排列——有时甚至是闭合的圆圈或环面。Chris Olah、Anthropic 的 Josh Batson 以及 Engels 等人，都在包括 ",[240,494,497],{"href":495,"rel":496},"https:\u002F\u002Farxiv.org\u002Fabs\u002F2405.14860",[244],"并非所有语言模型特征都是一维线性的"," 在内的研究中展示了这一现象。",[11,500,501],{},"这些发现引人入胜，却也留下了一个令人不安的空白：我们看到了流形，却无法精确地说出这个流形与它所表征的「概念」之间究竟是什么关系。为什么「年份」这样一个本该是直线的东西，会在模型内部被扭曲成一条在高维空间中蜿蜒盘绕的曲线？「颜色」排列成圆形是巧合还是必然？向量之间的余弦相似度，真的能告诉我们两个概念在语义上有多接近吗？",[11,503,504,505,510],{},"三位数学家——分别来自帝国理工学院、爱丁堡大学和布里斯托大学的 Alexander Modell、Patrick Rubin-Delanchy 与 Nick Whiteley——在他们的论文 ",[240,506,509],{"href":507,"rel":508},"https:\u002F\u002Farxiv.org\u002Fpdf\u002F2505.18235",[244],"《大语言模型中表征流形的起源》"," 中，首次尝试为这些问题提供一个「最小可用」的数学理论。本文梳理他们论证的脉络，并给出我个人的一些思考。",[224,512,513],{"id":513},"线性表征假说",[11,515,516],{},"机制可解释性领域长期以来由一个核心信念所锚定，即线性表征假说（LRH）：模型将人类可解释的「特征」——例如「长着毛茸茸的耳朵」「提到了埃菲尔铁塔」「使用阿拉伯语」——编码为表征空间中一组近似正交的方向向量。给定输入的表示，则是这些方向向量的稀疏线性组合，权重由各特征是否存在及其程度决定。如今被广泛采用的可解释性工具稀疏自编码器（SAE），正是直接建立在这一假说之上：训练一个带有稀疏性惩罚的自编码器，去逼近这些「字典向量」。",[11,518,519],{},"然而，越来越多的证据表明，这种「非黑即白、一特征一方向」的模型无法解释某些现象。作者引用了大量的文献：模加任务中的数字被编码为圆；多语言模型中出现了环状结构；「日期」与「星期几」呈现出扭曲的环面几何；甚至在模拟的隐马尔可夫模型中涌现出分形几何。这些例子有一个共同点：特征不再是一条直线，而是一个完整、连续、可能非线性弯曲的流形。",[11,521,522],{},"因此，该领域提出了 LRH 的推广版本——多维线性表征假说：",[24,524,527],{"className":525,"translate":28},[526],"katex-display",[24,528,530,612],{"className":529,"translate":28},[27],[24,531,533],{"className":532},[32],[34,534,536],{"xmlns":36,"display":535},"block",[38,537,538,609],{},[41,539,540,545,550,553,556,559,582,590,592,594,596,603,605,607],{},[541,542,544],"mi",{"mathvariant":543},"normal","Ψ",[546,547,549],"mo",{"stretchy":548},"false","(",[541,551,552],{},"x",[546,554,555],{"stretchy":548},")",[546,557,558],{},"=",[560,561,562,565],"munder",{},[546,563,564],{},"∑",[41,566,567,570,573,576,578,580],{},[541,568,569],{},"f",[546,571,572],{},"∈",[541,574,575],{},"F",[546,577,549],{"stretchy":548},[541,579,552],{},[546,581,555],{"stretchy":548},[583,584,585,588],"msub",{},[541,586,587],{},"ρ",[541,589,569],{},[546,591,549],{"stretchy":548},[541,593,552],{},[546,595,555],{"stretchy":548},[583,597,598,601],{},[541,599,600],{},"v",[541,602,569],{},[546,604,549],{"stretchy":548},[541,606,552],{},[546,608,555],{"stretchy":548},[48,610,611],{"encoding":50},"\\Psi(x) = \\sum_{f \\in F(x)} \\rho_f(x) v_f(x)",[24,613,615,648],{"className":614,"ariaHidden":56},[55],[24,616,618,622,625,629,633,637,641,645],{"className":617},[60],[24,619],{"className":620,"style":621},[64],"height:1em;vertical-align:-0.25em;",[24,623,544],{"className":624},[69],[24,626,549],{"className":627},[628],"mopen",[24,630,552],{"className":631},[69,632],"mathnormal",[24,634,555],{"className":635},[636],"mclose",[24,638],{"className":639,"style":640},[79],"margin-right:0.2778em;",[24,642,558],{"className":643},[644],"mrel",[24,646],{"className":647,"style":640},[79],[24,649,651,655,728,732,776,779,782,785,827,830,833],{"className":650},[60],[24,652],{"className":653,"style":654},[64],"height:2.566em;vertical-align:-1.516em;",[24,656,660],{"className":657},[658,659],"mop","op-limits",[24,661,663,719],{"className":662},[84,131],[24,664,666,716],{"className":665},[88],[24,667,670,703],{"className":668,"style":669},[92],"height:1.05em;",[24,671,673,677],{"style":672},"top:-1.809em;margin-left:0em;",[24,674],{"className":675,"style":676},[100],"height:3.05em;",[24,678,680],{"className":679},[109,110,111,108],[24,681,683,687,690,694,697,700],{"className":682},[69,108],[24,684,569],{"className":685,"style":686},[69,632,108],"margin-right:0.1076em;",[24,688,572],{"className":689},[644,108],[24,691,575],{"className":692,"style":693},[69,632,108],"margin-right:0.1389em;",[24,695,549],{"className":696},[628,108],[24,698,552],{"className":699},[69,632,108],[24,701,555],{"className":702},[636,108],[24,704,706,709],{"style":705},"top:-3.05em;",[24,707],{"className":708,"style":676},[100],[24,710,711],{},[24,712,564],{"className":713},[658,714,715],"op-symbol","large-op",[24,717,157],{"className":718},[156],[24,720,722],{"className":721},[88],[24,723,726],{"className":724,"style":725},[92],"height:1.516em;",[24,727],{},[24,729],{"className":730,"style":731},[79],"margin-right:0.1667em;",[24,733,735,738],{"className":734},[69],[24,736,587],{"className":737},[69,632],[24,739,742],{"className":740},[741],"msupsub",[24,743,745,767],{"className":744},[84,131],[24,746,748,764],{"className":747},[88],[24,749,752],{"className":750,"style":751},[92],"height:0.3361em;",[24,753,755,758],{"style":754},"top:-2.55em;margin-left:0em;margin-right:0.05em;",[24,756],{"className":757,"style":101},[100],[24,759,761],{"className":760},[109,110,111,108],[24,762,569],{"className":763,"style":686},[69,632,108],[24,765,157],{"className":766},[156],[24,768,770],{"className":769},[88],[24,771,774],{"className":772,"style":773},[92],"height:0.2861em;",[24,775],{},[24,777,549],{"className":778},[628],[24,780,552],{"className":781},[69,632],[24,783,555],{"className":784},[636],[24,786,788,792],{"className":787},[69],[24,789,600],{"className":790,"style":791},[69,632],"margin-right:0.0359em;",[24,793,795],{"className":794},[741],[24,796,798,819],{"className":797},[84,131],[24,799,801,816],{"className":800},[88],[24,802,804],{"className":803,"style":751},[92],[24,805,807,810],{"style":806},"top:-2.55em;margin-left:-0.0359em;margin-right:0.05em;",[24,808],{"className":809,"style":101},[100],[24,811,813],{"className":812},[109,110,111,108],[24,814,569],{"className":815,"style":686},[69,632,108],[24,817,157],{"className":818},[156],[24,820,822],{"className":821},[88],[24,823,825],{"className":824,"style":773},[92],[24,826],{},[24,828,549],{"className":829},[628],[24,831,552],{"className":832},[69,632],[24,834,555],{"className":835},[636],[11,837,838,839,925,926,1000,1001,1030,1031,1115,1116,1185],{},"其中 ",[24,840,842,866],{"className":841,"translate":28},[27],[24,843,845],{"className":844},[32],[34,846,847],{"xmlns":36},[38,848,849,863],{},[41,850,851,857,859,861],{},[583,852,853,855],{},[541,854,600],{},[541,856,569],{},[546,858,549],{"stretchy":548},[541,860,552],{},[546,862,555],{"stretchy":548},[48,864,865],{"encoding":50},"v_f(x)",[24,867,869],{"className":868,"ariaHidden":56},[55],[24,870,872,876,916,919,922],{"className":871},[60],[24,873],{"className":874,"style":875},[64],"height:1.0361em;vertical-align:-0.2861em;",[24,877,879,882],{"className":878},[69],[24,880,600],{"className":881,"style":791},[69,632],[24,883,885],{"className":884},[741],[24,886,888,908],{"className":887},[84,131],[24,889,891,905],{"className":890},[88],[24,892,894],{"className":893,"style":751},[92],[24,895,896,899],{"style":806},[24,897],{"className":898,"style":101},[100],[24,900,902],{"className":901},[109,110,111,108],[24,903,569],{"className":904,"style":686},[69,632,108],[24,906,157],{"className":907},[156],[24,909,911],{"className":910},[88],[24,912,914],{"className":913,"style":773},[92],[24,915],{},[24,917,549],{"className":918},[628],[24,920,552],{"className":921},[69,632],[24,923,555],{"className":924},[636]," 不再是固定方向，而是可以在某个子空间 ",[24,927,929,948],{"className":928,"translate":28},[27],[24,930,932],{"className":931},[32],[34,933,934],{"xmlns":36},[38,935,936,945],{},[41,937,938],{},[583,939,940,943],{},[541,941,942],{},"V",[541,944,569],{},[48,946,947],{"encoding":50},"V_f",[24,949,951],{"className":950,"ariaHidden":56},[55],[24,952,954,958],{"className":953},[60],[24,955],{"className":956,"style":957},[64],"height:0.9694em;vertical-align:-0.2861em;",[24,959,961,965],{"className":960},[69],[24,962,942],{"className":963,"style":964},[69,632],"margin-right:0.2222em;",[24,966,968],{"className":967},[741],[24,969,971,992],{"className":970},[84,131],[24,972,974,989],{"className":973},[88],[24,975,977],{"className":976,"style":751},[92],[24,978,980,983],{"style":979},"top:-2.55em;margin-left:-0.2222em;margin-right:0.05em;",[24,981],{"className":982,"style":101},[100],[24,984,986],{"className":985},[109,110,111,108],[24,987,569],{"className":988,"style":686},[69,632,108],[24,990,157],{"className":991},[156],[24,993,995],{"className":994},[88],[24,996,998],{"className":997,"style":773},[92],[24,999],{}," 内随输入 ",[24,1002,1004,1017],{"className":1003,"translate":28},[27],[24,1005,1007],{"className":1006},[32],[34,1008,1009],{"xmlns":36},[38,1010,1011,1015],{},[41,1012,1013],{},[541,1014,552],{},[48,1016,552],{"encoding":50},[24,1018,1020],{"className":1019,"ariaHidden":56},[55],[24,1021,1023,1027],{"className":1022},[60],[24,1024],{"className":1025,"style":1026},[64],"height:0.4306em;",[24,1028,552],{"className":1029},[69,632]," 连续变化。标准的 LRH 只是 ",[24,1032,1034,1057],{"className":1033,"translate":28},[27],[24,1035,1037],{"className":1036},[32],[34,1038,1039],{"xmlns":36},[38,1040,1041,1055],{},[41,1042,1043,1049,1051,1053],{},[583,1044,1045,1047],{},[541,1046,600],{},[541,1048,569],{},[546,1050,549],{"stretchy":548},[541,1052,552],{},[546,1054,555],{"stretchy":548},[48,1056,865],{"encoding":50},[24,1058,1060],{"className":1059,"ariaHidden":56},[55],[24,1061,1063,1066,1106,1109,1112],{"className":1062},[60],[24,1064],{"className":1065,"style":875},[64],[24,1067,1069,1072],{"className":1068},[69],[24,1070,600],{"className":1071,"style":791},[69,632],[24,1073,1075],{"className":1074},[741],[24,1076,1078,1098],{"className":1077},[84,131],[24,1079,1081,1095],{"className":1080},[88],[24,1082,1084],{"className":1083,"style":751},[92],[24,1085,1086,1089],{"style":806},[24,1087],{"className":1088,"style":101},[100],[24,1090,1092],{"className":1091},[109,110,111,108],[24,1093,569],{"className":1094,"style":686},[69,632,108],[24,1096,157],{"className":1097},[156],[24,1099,1101],{"className":1100},[88],[24,1102,1104],{"className":1103,"style":773},[92],[24,1105],{},[24,1107,549],{"className":1108},[628],[24,1110,552],{"className":1111},[69,632],[24,1113,555],{"className":1114},[636]," 恒定且 ",[24,1117,1119,1136],{"className":1118,"translate":28},[27],[24,1120,1122],{"className":1121},[32],[34,1123,1124],{"xmlns":36},[38,1125,1126,1134],{},[41,1127,1128],{},[583,1129,1130,1132],{},[541,1131,942],{},[541,1133,569],{},[48,1135,947],{"encoding":50},[24,1137,1139],{"className":1138,"ariaHidden":56},[55],[24,1140,1142,1145],{"className":1141},[60],[24,1143],{"className":1144,"style":957},[64],[24,1146,1148,1151],{"className":1147},[69],[24,1149,942],{"className":1150,"style":964},[69,632],[24,1152,1154],{"className":1153},[741],[24,1155,1157,1177],{"className":1156},[84,131],[24,1158,1160,1174],{"className":1159},[88],[24,1161,1163],{"className":1162,"style":751},[92],[24,1164,1165,1168],{"style":979},[24,1166],{"className":1167,"style":101},[100],[24,1169,1171],{"className":1170},[109,110,111,108],[24,1172,569],{"className":1173,"style":686},[69,632,108],[24,1175,157],{"className":1176},[156],[24,1178,1180],{"className":1179},[88],[24,1181,1183],{"className":1182,"style":773},[92],[24,1184],{}," 一维的特例。",[11,1187,1188,1189,1218,1219,1288],{},"这篇论文想要处理的，恰恰是这个推广假说中最核心也最含糊的部分：特征 ",[24,1190,1192,1205],{"className":1191,"translate":28},[27],[24,1193,1195],{"className":1194},[32],[34,1196,1197],{"xmlns":36},[38,1198,1199,1203],{},[41,1200,1201],{},[541,1202,569],{},[48,1204,569],{"encoding":50},[24,1206,1208],{"className":1207,"ariaHidden":56},[55],[24,1209,1211,1215],{"className":1210},[60],[24,1212],{"className":1213,"style":1214},[64],"height:0.8889em;vertical-align:-0.1944em;",[24,1216,569],{"className":1217,"style":686},[69,632]," 在子空间 ",[24,1220,1222,1239],{"className":1221,"translate":28},[27],[24,1223,1225],{"className":1224},[32],[34,1226,1227],{"xmlns":36},[38,1228,1229,1237],{},[41,1230,1231],{},[583,1232,1233,1235],{},[541,1234,942],{},[541,1236,569],{},[48,1238,947],{"encoding":50},[24,1240,1242],{"className":1241,"ariaHidden":56},[55],[24,1243,1245,1248],{"className":1244},[60],[24,1246],{"className":1247,"style":957},[64],[24,1249,1251,1254],{"className":1250},[69],[24,1252,942],{"className":1253,"style":964},[69,632],[24,1255,1257],{"className":1256},[741],[24,1258,1260,1280],{"className":1259},[84,131],[24,1261,1263,1277],{"className":1262},[88],[24,1264,1266],{"className":1265,"style":751},[92],[24,1267,1268,1271],{"style":979},[24,1269],{"className":1270,"style":101},[100],[24,1272,1274],{"className":1273},[109,110,111,108],[24,1275,569],{"className":1276,"style":686},[69,632,108],[24,1278,157],{"className":1279},[156],[24,1281,1283],{"className":1282},[88],[24,1284,1286],{"className":1285,"style":773},[92],[24,1287],{}," 中以何种流形形态呈现，这个流形与「特征」本身之间又是什么关系？",[224,1290,1292],{"id":1291},"特征即度量空间","「特征」即度量空间",[11,1294,1295],{},"论文中最优雅的一步，是先解决一个听起来有些哲学、实际上至关重要的问题：「特征」究竟是什么？",[11,1297,1298,1299,1439],{},"作者给出的答案出人意料地简洁：特征是一个度量空间 ",[24,1300,1302,1335],{"className":1301,"translate":28},[27],[24,1303,1305],{"className":1304},[32],[34,1306,1307],{"xmlns":36},[38,1308,1309,1332],{},[41,1310,1311,1313,1320,1323,1330],{},[546,1312,549],{"stretchy":548},[583,1314,1315,1318],{},[541,1316,1317],{},"Z",[541,1319,569],{},[546,1321,1322],{"separator":56},",",[583,1324,1325,1328],{},[541,1326,1327],{},"d",[541,1329,569],{},[546,1331,555],{"stretchy":548},[48,1333,1334],{"encoding":50},"(Z_f, d_f)",[24,1336,1338],{"className":1337,"ariaHidden":56},[55],[24,1339,1341,1344,1347,1389,1393,1396,1436],{"className":1340},[60],[24,1342],{"className":1343,"style":875},[64],[24,1345,549],{"className":1346},[628],[24,1348,1350,1354],{"className":1349},[69],[24,1351,1317],{"className":1352,"style":1353},[69,632],"margin-right:0.0715em;",[24,1355,1357],{"className":1356},[741],[24,1358,1360,1381],{"className":1359},[84,131],[24,1361,1363,1378],{"className":1362},[88],[24,1364,1366],{"className":1365,"style":751},[92],[24,1367,1369,1372],{"style":1368},"top:-2.55em;margin-left:-0.0715em;margin-right:0.05em;",[24,1370],{"className":1371,"style":101},[100],[24,1373,1375],{"className":1374},[109,110,111,108],[24,1376,569],{"className":1377,"style":686},[69,632,108],[24,1379,157],{"className":1380},[156],[24,1382,1384],{"className":1383},[88],[24,1385,1387],{"className":1386,"style":773},[92],[24,1388],{},[24,1390,1322],{"className":1391},[1392],"mpunct",[24,1394],{"className":1395,"style":731},[79],[24,1397,1399,1402],{"className":1398},[69],[24,1400,1327],{"className":1401},[69,632],[24,1403,1405],{"className":1404},[741],[24,1406,1408,1428],{"className":1407},[84,131],[24,1409,1411,1425],{"className":1410},[88],[24,1412,1414],{"className":1413,"style":751},[92],[24,1415,1416,1419],{"style":754},[24,1417],{"className":1418,"style":101},[100],[24,1420,1422],{"className":1421},[109,110,111,108],[24,1423,569],{"className":1424,"style":686},[69,632,108],[24,1426,157],{"className":1427},[156],[24,1429,1431],{"className":1430},[88],[24,1432,1434],{"className":1433,"style":773},[92],[24,1435],{},[24,1437,555],{"className":1438},[636],"——一个集合连同定义在该集合上的某种「距离」概念。这个定义看似平淡，实则表现力极强：",[1441,1442,1443,1518,1663],"ul",{},[1444,1445,1446,1447,1517],"li",{},"原子特征（是否有猫）：",[24,1448,1450,1468],{"className":1449,"translate":28},[27],[24,1451,1453],{"className":1452},[32],[34,1454,1455],{"xmlns":36},[38,1456,1457,1465],{},[41,1458,1459],{},[583,1460,1461,1463],{},[541,1462,1317],{},[541,1464,569],{},[48,1466,1467],{"encoding":50},"Z_f",[24,1469,1471],{"className":1470,"ariaHidden":56},[55],[24,1472,1474,1477],{"className":1473},[60],[24,1475],{"className":1476,"style":957},[64],[24,1478,1480,1483],{"className":1479},[69],[24,1481,1317],{"className":1482,"style":1353},[69,632],[24,1484,1486],{"className":1485},[741],[24,1487,1489,1509],{"className":1488},[84,131],[24,1490,1492,1506],{"className":1491},[88],[24,1493,1495],{"className":1494,"style":751},[92],[24,1496,1497,1500],{"style":1368},[24,1498],{"className":1499,"style":101},[100],[24,1501,1503],{"className":1502},[109,110,111,108],[24,1504,569],{"className":1505,"style":686},[69,632,108],[24,1507,157],{"className":1508},[156],[24,1510,1512],{"className":1511},[88],[24,1513,1515],{"className":1514,"style":773},[92],[24,1516],{}," 是单元素集合；",[1444,1519,1520,1521,1590,1591,1662],{},"层次特征（动物的分类学树）：",[24,1522,1524,1541],{"className":1523,"translate":28},[27],[24,1525,1527],{"className":1526},[32],[34,1528,1529],{"xmlns":36},[38,1530,1531,1539],{},[41,1532,1533],{},[583,1534,1535,1537],{},[541,1536,1317],{},[541,1538,569],{},[48,1540,1467],{"encoding":50},[24,1542,1544],{"className":1543,"ariaHidden":56},[55],[24,1545,1547,1550],{"className":1546},[60],[24,1548],{"className":1549,"style":957},[64],[24,1551,1553,1556],{"className":1552},[69],[24,1554,1317],{"className":1555,"style":1353},[69,632],[24,1557,1559],{"className":1558},[741],[24,1560,1562,1582],{"className":1561},[84,131],[24,1563,1565,1579],{"className":1564},[88],[24,1566,1568],{"className":1567,"style":751},[92],[24,1569,1570,1573],{"style":1368},[24,1571],{"className":1572,"style":101},[100],[24,1574,1576],{"className":1575},[109,110,111,108],[24,1577,569],{"className":1578,"style":686},[69,632,108],[24,1580,157],{"className":1581},[156],[24,1583,1585],{"className":1584},[88],[24,1586,1588],{"className":1587,"style":773},[92],[24,1589],{}," 是离散集合，",[24,1592,1594,1612],{"className":1593,"translate":28},[27],[24,1595,1597],{"className":1596},[32],[34,1598,1599],{"xmlns":36},[38,1600,1601,1609],{},[41,1602,1603],{},[583,1604,1605,1607],{},[541,1606,1327],{},[541,1608,569],{},[48,1610,1611],{"encoding":50},"d_f",[24,1613,1615],{"className":1614,"ariaHidden":56},[55],[24,1616,1618,1622],{"className":1617},[60],[24,1619],{"className":1620,"style":1621},[64],"height:0.9805em;vertical-align:-0.2861em;",[24,1623,1625,1628],{"className":1624},[69],[24,1626,1327],{"className":1627},[69,632],[24,1629,1631],{"className":1630},[741],[24,1632,1634,1654],{"className":1633},[84,131],[24,1635,1637,1651],{"className":1636},[88],[24,1638,1640],{"className":1639,"style":751},[92],[24,1641,1642,1645],{"style":754},[24,1643],{"className":1644,"style":101},[100],[24,1646,1648],{"className":1647},[109,110,111,108],[24,1649,569],{"className":1650,"style":686},[69,632,108],[24,1652,157],{"className":1653},[156],[24,1655,1657],{"className":1656},[88],[24,1658,1660],{"className":1659,"style":773},[92],[24,1661],{}," 是其上的树距离；",[1444,1664,1665,1666,1735],{},"连续特征（颜色色相、一年中的第几天、日历年）：",[24,1667,1669,1686],{"className":1668,"translate":28},[27],[24,1670,1672],{"className":1671},[32],[34,1673,1674],{"xmlns":36},[38,1675,1676,1684],{},[41,1677,1678],{},[583,1679,1680,1682],{},[541,1681,1317],{},[541,1683,569],{},[48,1685,1467],{"encoding":50},[24,1687,1689],{"className":1688,"ariaHidden":56},[55],[24,1690,1692,1695],{"className":1691},[60],[24,1693],{"className":1694,"style":957},[64],[24,1696,1698,1701],{"className":1697},[69],[24,1699,1317],{"className":1700,"style":1353},[69,632],[24,1702,1704],{"className":1703},[741],[24,1705,1707,1727],{"className":1706},[84,131],[24,1708,1710,1724],{"className":1709},[88],[24,1711,1713],{"className":1712,"style":751},[92],[24,1714,1715,1718],{"style":1368},[24,1716],{"className":1717,"style":101},[100],[24,1719,1721],{"className":1720},[109,110,111,108],[24,1722,569],{"className":1723,"style":686},[69,632,108],[24,1725,157],{"className":1726},[156],[24,1728,1730],{"className":1729},[88],[24,1731,1733],{"className":1732,"style":773},[92],[24,1734],{}," 可以是区间、圆或更高维的欧几里得空间。",[11,1737,1738],{},"这一框架比学习理论中常见的「欧几里得空间」或「超球面」假设更灵活，又比解耦文献中带群结构的黎曼流形简单得多、也更易于处理。这是典型的数学家式选择：在表达力与可处理性之间取得恰当的平衡。",[11,1740,1741],{},"在「特征即度量空间」这一定义就位后，作者提出了论文的第一个核心假说：",[356,1743,1744],{},[11,1745,1746,1747,1835,1836,1920,1921,2135,2136,442],{},"假说 1（连续对应假说）。特征值 ",[24,1748,1750,1775],{"className":1749,"translate":28},[27],[24,1751,1753],{"className":1752},[32],[34,1754,1755],{"xmlns":36},[38,1756,1757,1772],{},[41,1758,1759,1766,1768,1770],{},[583,1760,1761,1764],{},[541,1762,1763],{},"z",[541,1765,569],{},[546,1767,549],{"stretchy":548},[541,1769,552],{},[546,1771,555],{"stretchy":548},[48,1773,1774],{"encoding":50},"z_f(x)",[24,1776,1778],{"className":1777,"ariaHidden":56},[55],[24,1779,1781,1784,1826,1829,1832],{"className":1780},[60],[24,1782],{"className":1783,"style":875},[64],[24,1785,1787,1791],{"className":1786},[69],[24,1788,1763],{"className":1789,"style":1790},[69,632],"margin-right:0.044em;",[24,1792,1794],{"className":1793},[741],[24,1795,1797,1818],{"className":1796},[84,131],[24,1798,1800,1815],{"className":1799},[88],[24,1801,1803],{"className":1802,"style":751},[92],[24,1804,1806,1809],{"style":1805},"top:-2.55em;margin-left:-0.044em;margin-right:0.05em;",[24,1807],{"className":1808,"style":101},[100],[24,1810,1812],{"className":1811},[109,110,111,108],[24,1813,569],{"className":1814,"style":686},[69,632,108],[24,1816,157],{"className":1817},[156],[24,1819,1821],{"className":1820},[88],[24,1822,1824],{"className":1823,"style":773},[92],[24,1825],{},[24,1827,549],{"className":1828},[628],[24,1830,552],{"className":1831},[69,632],[24,1833,555],{"className":1834},[636]," 与表示方向 ",[24,1837,1839,1862],{"className":1838,"translate":28},[27],[24,1840,1842],{"className":1841},[32],[34,1843,1844],{"xmlns":36},[38,1845,1846,1860],{},[41,1847,1848,1854,1856,1858],{},[583,1849,1850,1852],{},[541,1851,600],{},[541,1853,569],{},[546,1855,549],{"stretchy":548},[541,1857,552],{},[546,1859,555],{"stretchy":548},[48,1861,865],{"encoding":50},[24,1863,1865],{"className":1864,"ariaHidden":56},[55],[24,1866,1868,1871,1911,1914,1917],{"className":1867},[60],[24,1869],{"className":1870,"style":875},[64],[24,1872,1874,1877],{"className":1873},[69],[24,1875,600],{"className":1876,"style":791},[69,632],[24,1878,1880],{"className":1879},[741],[24,1881,1883,1903],{"className":1882},[84,131],[24,1884,1886,1900],{"className":1885},[88],[24,1887,1889],{"className":1888,"style":751},[92],[24,1890,1891,1894],{"style":806},[24,1892],{"className":1893,"style":101},[100],[24,1895,1897],{"className":1896},[109,110,111,108],[24,1898,569],{"className":1899,"style":686},[69,632,108],[24,1901,157],{"className":1902},[156],[24,1904,1906],{"className":1905},[88],[24,1907,1909],{"className":1908,"style":773},[92],[24,1910],{},[24,1912,549],{"className":1913},[628],[24,1915,552],{"className":1916},[69,632],[24,1918,555],{"className":1919},[636]," 之间存在连续、可逆的一一对应；即存在连续映射 ",[24,1922,1924,1973],{"className":1923,"translate":28},[27],[24,1925,1927],{"className":1926},[32],[34,1928,1929],{"xmlns":36},[38,1930,1931,1970],{},[41,1932,1933,1940,1943,1949,1952],{},[583,1934,1935,1938],{},[541,1936,1937],{},"ϕ",[541,1939,569],{},[546,1941,1942],{},":",[583,1944,1945,1947],{},[541,1946,1317],{},[541,1948,569],{},[546,1950,1951],{},"→",[1953,1954,1955,1958],"msup",{},[541,1956,1957],{},"S",[41,1959,1960,1963,1966],{},[541,1961,1962],{},"D",[546,1964,1965],{},"−",[1967,1968,1969],"mn",{},"1",[48,1971,1972],{"encoding":50},"\\phi_f: Z_f \\to S^{D-1}",[24,1974,1976,2031,2086],{"className":1975,"ariaHidden":56},[55],[24,1977,1979,1982,2022,2025,2028],{"className":1978},[60],[24,1980],{"className":1981,"style":1621},[64],[24,1983,1985,1988],{"className":1984},[69],[24,1986,1937],{"className":1987},[69,632],[24,1989,1991],{"className":1990},[741],[24,1992,1994,2014],{"className":1993},[84,131],[24,1995,1997,2011],{"className":1996},[88],[24,1998,2000],{"className":1999,"style":751},[92],[24,2001,2002,2005],{"style":754},[24,2003],{"className":2004,"style":101},[100],[24,2006,2008],{"className":2007},[109,110,111,108],[24,2009,569],{"className":2010,"style":686},[69,632,108],[24,2012,157],{"className":2013},[156],[24,2015,2017],{"className":2016},[88],[24,2018,2020],{"className":2019,"style":773},[92],[24,2021],{},[24,2023],{"className":2024,"style":640},[79],[24,2026,1942],{"className":2027},[644],[24,2029],{"className":2030,"style":640},[79],[24,2032,2034,2037,2077,2080,2083],{"className":2033},[60],[24,2035],{"className":2036,"style":957},[64],[24,2038,2040,2043],{"className":2039},[69],[24,2041,1317],{"className":2042,"style":1353},[69,632],[24,2044,2046],{"className":2045},[741],[24,2047,2049,2069],{"className":2048},[84,131],[24,2050,2052,2066],{"className":2051},[88],[24,2053,2055],{"className":2054,"style":751},[92],[24,2056,2057,2060],{"style":1368},[24,2058],{"className":2059,"style":101},[100],[24,2061,2063],{"className":2062},[109,110,111,108],[24,2064,569],{"className":2065,"style":686},[69,632,108],[24,2067,157],{"className":2068},[156],[24,2070,2072],{"className":2071},[88],[24,2073,2075],{"className":2074,"style":773},[92],[24,2076],{},[24,2078],{"className":2079,"style":640},[79],[24,2081,1951],{"className":2082},[644],[24,2084],{"className":2085,"style":640},[79],[24,2087,2089,2093],{"className":2088},[60],[24,2090],{"className":2091,"style":2092},[64],"height:0.8413em;",[24,2094,2096,2100],{"className":2095},[69],[24,2097,1957],{"className":2098,"style":2099},[69,632],"margin-right:0.0576em;",[24,2101,2103],{"className":2102},[741],[24,2104,2106],{"className":2105},[84],[24,2107,2109],{"className":2108},[88],[24,2110,2112],{"className":2111,"style":2092},[92],[24,2113,2115,2118],{"style":2114},"top:-3.063em;margin-right:0.05em;",[24,2116],{"className":2117,"style":101},[100],[24,2119,2121],{"className":2120},[109,110,111,108],[24,2122,2124,2128,2132],{"className":2123},[69,108],[24,2125,1962],{"className":2126,"style":2127},[69,632,108],"margin-right:0.0278em;",[24,2129,1965],{"className":2130},[2131,108],"mbin",[24,2133,1969],{"className":2134},[69,108],"（映到单位超球面），使得 ",[24,2137,2139,2187],{"className":2138,"translate":28},[27],[24,2140,2142],{"className":2141},[32],[34,2143,2144],{"xmlns":36},[38,2145,2146,2184],{},[41,2147,2148,2154,2156,2158,2160,2162,2168,2170,2176,2178,2180,2182],{},[583,2149,2150,2152],{},[541,2151,600],{},[541,2153,569],{},[546,2155,549],{"stretchy":548},[541,2157,552],{},[546,2159,555],{"stretchy":548},[546,2161,558],{},[583,2163,2164,2166],{},[541,2165,1937],{},[541,2167,569],{},[546,2169,549],{"stretchy":548},[583,2171,2172,2174],{},[541,2173,1763],{},[541,2175,569],{},[546,2177,549],{"stretchy":548},[541,2179,552],{},[546,2181,555],{"stretchy":548},[546,2183,555],{"stretchy":548},[48,2185,2186],{"encoding":50},"v_f(x) = \\phi_f(z_f(x))",[24,2188,2190,2254],{"className":2189,"ariaHidden":56},[55],[24,2191,2193,2196,2236,2239,2242,2245,2248,2251],{"className":2192},[60],[24,2194],{"className":2195,"style":875},[64],[24,2197,2199,2202],{"className":2198},[69],[24,2200,600],{"className":2201,"style":791},[69,632],[24,2203,2205],{"className":2204},[741],[24,2206,2208,2228],{"className":2207},[84,131],[24,2209,2211,2225],{"className":2210},[88],[24,2212,2214],{"className":2213,"style":751},[92],[24,2215,2216,2219],{"style":806},[24,2217],{"className":2218,"style":101},[100],[24,2220,2222],{"className":2221},[109,110,111,108],[24,2223,569],{"className":2224,"style":686},[69,632,108],[24,2226,157],{"className":2227},[156],[24,2229,2231],{"className":2230},[88],[24,2232,2234],{"className":2233,"style":773},[92],[24,2235],{},[24,2237,549],{"className":2238},[628],[24,2240,552],{"className":2241},[69,632],[24,2243,555],{"className":2244},[636],[24,2246],{"className":2247,"style":640},[79],[24,2249,558],{"className":2250},[644],[24,2252],{"className":2253,"style":640},[79],[24,2255,2257,2260,2300,2303,2343,2346,2349],{"className":2256},[60],[24,2258],{"className":2259,"style":875},[64],[24,2261,2263,2266],{"className":2262},[69],[24,2264,1937],{"className":2265},[69,632],[24,2267,2269],{"className":2268},[741],[24,2270,2272,2292],{"className":2271},[84,131],[24,2273,2275,2289],{"className":2274},[88],[24,2276,2278],{"className":2277,"style":751},[92],[24,2279,2280,2283],{"style":754},[24,2281],{"className":2282,"style":101},[100],[24,2284,2286],{"className":2285},[109,110,111,108],[24,2287,569],{"className":2288,"style":686},[69,632,108],[24,2290,157],{"className":2291},[156],[24,2293,2295],{"className":2294},[88],[24,2296,2298],{"className":2297,"style":773},[92],[24,2299],{},[24,2301,549],{"className":2302},[628],[24,2304,2306,2309],{"className":2305},[69],[24,2307,1763],{"className":2308,"style":1790},[69,632],[24,2310,2312],{"className":2311},[741],[24,2313,2315,2335],{"className":2314},[84,131],[24,2316,2318,2332],{"className":2317},[88],[24,2319,2321],{"className":2320,"style":751},[92],[24,2322,2323,2326],{"style":1805},[24,2324],{"className":2325,"style":101},[100],[24,2327,2329],{"className":2328},[109,110,111,108],[24,2330,569],{"className":2331,"style":686},[69,632,108],[24,2333,157],{"className":2334},[156],[24,2336,2338],{"className":2337},[88],[24,2339,2341],{"className":2340,"style":773},[92],[24,2342],{},[24,2344,549],{"className":2345},[628],[24,2347,552],{"className":2348},[69,632],[24,2350,2352],{"className":2351},[636],"))",[11,2354,2355,2356,2425,2426,2496,2497,2570,2571,2640,2641,2710,2711,2780,2781,2850,2851,2920,2921,2990],{},"结合 ",[24,2357,2359,2376],{"className":2358,"translate":28},[27],[24,2360,2362],{"className":2361},[32],[34,2363,2364],{"xmlns":36},[38,2365,2366,2374],{},[41,2367,2368],{},[583,2369,2370,2372],{},[541,2371,1317],{},[541,2373,569],{},[48,2375,1467],{"encoding":50},[24,2377,2379],{"className":2378,"ariaHidden":56},[55],[24,2380,2382,2385],{"className":2381},[60],[24,2383],{"className":2384,"style":957},[64],[24,2386,2388,2391],{"className":2387},[69],[24,2389,1317],{"className":2390,"style":1353},[69,632],[24,2392,2394],{"className":2393},[741],[24,2395,2397,2417],{"className":2396},[84,131],[24,2398,2400,2414],{"className":2399},[88],[24,2401,2403],{"className":2402,"style":751},[92],[24,2404,2405,2408],{"style":1368},[24,2406],{"className":2407,"style":101},[100],[24,2409,2411],{"className":2410},[109,110,111,108],[24,2412,569],{"className":2413,"style":686},[69,632,108],[24,2415,157],{"className":2416},[156],[24,2418,2420],{"className":2419},[88],[24,2421,2423],{"className":2422,"style":773},[92],[24,2424],{}," 是紧空间的这一技术前提，该假说立即导出一个干净的推论（命题 1）：",[24,2427,2429,2447],{"className":2428,"translate":28},[27],[24,2430,2432],{"className":2431},[32],[34,2433,2434],{"xmlns":36},[38,2435,2436,2444],{},[41,2437,2438],{},[583,2439,2440,2442],{},[541,2441,1937],{},[541,2443,569],{},[48,2445,2446],{"encoding":50},"\\phi_f",[24,2448,2450],{"className":2449,"ariaHidden":56},[55],[24,2451,2453,2456],{"className":2452},[60],[24,2454],{"className":2455,"style":1621},[64],[24,2457,2459,2462],{"className":2458},[69],[24,2460,1937],{"className":2461},[69,632],[24,2463,2465],{"className":2464},[741],[24,2466,2468,2488],{"className":2467},[84,131],[24,2469,2471,2485],{"className":2470},[88],[24,2472,2474],{"className":2473,"style":751},[92],[24,2475,2476,2479],{"style":754},[24,2477],{"className":2478,"style":101},[100],[24,2480,2482],{"className":2481},[109,110,111,108],[24,2483,569],{"className":2484,"style":686},[69,632,108],[24,2486,157],{"className":2487},[156],[24,2489,2491],{"className":2490},[88],[24,2492,2494],{"className":2493,"style":773},[92],[24,2495],{}," 是一个同胚。换言之，表征流形 ",[24,2498,2500,2519],{"className":2499,"translate":28},[27],[24,2501,2503],{"className":2502},[32],[34,2504,2505],{"xmlns":36},[38,2506,2507,2516],{},[41,2508,2509],{},[583,2510,2511,2514],{},[541,2512,2513],{},"M",[541,2515,569],{},[48,2517,2518],{"encoding":50},"M_f",[24,2520,2522],{"className":2521,"ariaHidden":56},[55],[24,2523,2525,2528],{"className":2524},[60],[24,2526],{"className":2527,"style":957},[64],[24,2529,2531,2535],{"className":2530},[69],[24,2532,2513],{"className":2533,"style":2534},[69,632],"margin-right:0.109em;",[24,2536,2538],{"className":2537},[741],[24,2539,2541,2562],{"className":2540},[84,131],[24,2542,2544,2559],{"className":2543},[88],[24,2545,2547],{"className":2546,"style":751},[92],[24,2548,2550,2553],{"style":2549},"top:-2.55em;margin-left:-0.109em;margin-right:0.05em;",[24,2551],{"className":2552,"style":101},[100],[24,2554,2556],{"className":2555},[109,110,111,108],[24,2557,569],{"className":2558,"style":686},[69,632,108],[24,2560,157],{"className":2561},[156],[24,2563,2565],{"className":2564},[88],[24,2566,2568],{"className":2567,"style":773},[92],[24,2569],{},"（即 ",[24,2572,2574,2591],{"className":2573,"translate":28},[27],[24,2575,2577],{"className":2576},[32],[34,2578,2579],{"xmlns":36},[38,2580,2581,2589],{},[41,2582,2583],{},[583,2584,2585,2587],{},[541,2586,1937],{},[541,2588,569],{},[48,2590,2446],{"encoding":50},[24,2592,2594],{"className":2593,"ariaHidden":56},[55],[24,2595,2597,2600],{"className":2596},[60],[24,2598],{"className":2599,"style":1621},[64],[24,2601,2603,2606],{"className":2602},[69],[24,2604,1937],{"className":2605},[69,632],[24,2607,2609],{"className":2608},[741],[24,2610,2612,2632],{"className":2611},[84,131],[24,2613,2615,2629],{"className":2614},[88],[24,2616,2618],{"className":2617,"style":751},[92],[24,2619,2620,2623],{"style":754},[24,2621],{"className":2622,"style":101},[100],[24,2624,2626],{"className":2625},[109,110,111,108],[24,2627,569],{"className":2628,"style":686},[69,632,108],[24,2630,157],{"className":2631},[156],[24,2633,2635],{"className":2634},[88],[24,2636,2638],{"className":2637,"style":773},[92],[24,2639],{}," 的像）与原始特征空间 ",[24,2642,2644,2661],{"className":2643,"translate":28},[27],[24,2645,2647],{"className":2646},[32],[34,2648,2649],{"xmlns":36},[38,2650,2651,2659],{},[41,2652,2653],{},[583,2654,2655,2657],{},[541,2656,1317],{},[541,2658,569],{},[48,2660,1467],{"encoding":50},[24,2662,2664],{"className":2663,"ariaHidden":56},[55],[24,2665,2667,2670],{"className":2666},[60],[24,2668],{"className":2669,"style":957},[64],[24,2671,2673,2676],{"className":2672},[69],[24,2674,1317],{"className":2675,"style":1353},[69,632],[24,2677,2679],{"className":2678},[741],[24,2680,2682,2702],{"className":2681},[84,131],[24,2683,2685,2699],{"className":2684},[88],[24,2686,2688],{"className":2687,"style":751},[92],[24,2689,2690,2693],{"style":1368},[24,2691],{"className":2692,"style":101},[100],[24,2694,2696],{"className":2695},[109,110,111,108],[24,2697,569],{"className":2698,"style":686},[69,632,108],[24,2700,157],{"className":2701},[156],[24,2703,2705],{"className":2704},[88],[24,2706,2708],{"className":2707,"style":773},[92],[24,2709],{}," 在拓扑上「形状相同」：若 ",[24,2712,2714,2731],{"className":2713,"translate":28},[27],[24,2715,2717],{"className":2716},[32],[34,2718,2719],{"xmlns":36},[38,2720,2721,2729],{},[41,2722,2723],{},[583,2724,2725,2727],{},[541,2726,1317],{},[541,2728,569],{},[48,2730,1467],{"encoding":50},[24,2732,2734],{"className":2733,"ariaHidden":56},[55],[24,2735,2737,2740],{"className":2736},[60],[24,2738],{"className":2739,"style":957},[64],[24,2741,2743,2746],{"className":2742},[69],[24,2744,1317],{"className":2745,"style":1353},[69,632],[24,2747,2749],{"className":2748},[741],[24,2750,2752,2772],{"className":2751},[84,131],[24,2753,2755,2769],{"className":2754},[88],[24,2756,2758],{"className":2757,"style":751},[92],[24,2759,2760,2763],{"style":1368},[24,2761],{"className":2762,"style":101},[100],[24,2764,2766],{"className":2765},[109,110,111,108],[24,2767,569],{"className":2768,"style":686},[69,632,108],[24,2770,157],{"className":2771},[156],[24,2773,2775],{"className":2774},[88],[24,2776,2778],{"className":2777,"style":773},[92],[24,2779],{}," 是区间，",[24,2782,2784,2801],{"className":2783,"translate":28},[27],[24,2785,2787],{"className":2786},[32],[34,2788,2789],{"xmlns":36},[38,2790,2791,2799],{},[41,2792,2793],{},[583,2794,2795,2797],{},[541,2796,2513],{},[541,2798,569],{},[48,2800,2518],{"encoding":50},[24,2802,2804],{"className":2803,"ariaHidden":56},[55],[24,2805,2807,2810],{"className":2806},[60],[24,2808],{"className":2809,"style":957},[64],[24,2811,2813,2816],{"className":2812},[69],[24,2814,2513],{"className":2815,"style":2534},[69,632],[24,2817,2819],{"className":2818},[741],[24,2820,2822,2842],{"className":2821},[84,131],[24,2823,2825,2839],{"className":2824},[88],[24,2826,2828],{"className":2827,"style":751},[92],[24,2829,2830,2833],{"style":2549},[24,2831],{"className":2832,"style":101},[100],[24,2834,2836],{"className":2835},[109,110,111,108],[24,2837,569],{"className":2838,"style":686},[69,632,108],[24,2840,157],{"className":2841},[156],[24,2843,2845],{"className":2844},[88],[24,2846,2848],{"className":2847,"style":773},[92],[24,2849],{}," 就是一条曲线；若 ",[24,2852,2854,2871],{"className":2853,"translate":28},[27],[24,2855,2857],{"className":2856},[32],[34,2858,2859],{"xmlns":36},[38,2860,2861,2869],{},[41,2862,2863],{},[583,2864,2865,2867],{},[541,2866,1317],{},[541,2868,569],{},[48,2870,1467],{"encoding":50},[24,2872,2874],{"className":2873,"ariaHidden":56},[55],[24,2875,2877,2880],{"className":2876},[60],[24,2878],{"className":2879,"style":957},[64],[24,2881,2883,2886],{"className":2882},[69],[24,2884,1317],{"className":2885,"style":1353},[69,632],[24,2887,2889],{"className":2888},[741],[24,2890,2892,2912],{"className":2891},[84,131],[24,2893,2895,2909],{"className":2894},[88],[24,2896,2898],{"className":2897,"style":751},[92],[24,2899,2900,2903],{"style":1368},[24,2901],{"className":2902,"style":101},[100],[24,2904,2906],{"className":2905},[109,110,111,108],[24,2907,569],{"className":2908,"style":686},[69,632,108],[24,2910,157],{"className":2911},[156],[24,2913,2915],{"className":2914},[88],[24,2916,2918],{"className":2917,"style":773},[92],[24,2919],{}," 是圆，",[24,2922,2924,2941],{"className":2923,"translate":28},[27],[24,2925,2927],{"className":2926},[32],[34,2928,2929],{"xmlns":36},[38,2930,2931,2939],{},[41,2932,2933],{},[583,2934,2935,2937],{},[541,2936,2513],{},[541,2938,569],{},[48,2940,2518],{"encoding":50},[24,2942,2944],{"className":2943,"ariaHidden":56},[55],[24,2945,2947,2950],{"className":2946},[60],[24,2948],{"className":2949,"style":957},[64],[24,2951,2953,2956],{"className":2952},[69],[24,2954,2513],{"className":2955,"style":2534},[69,632],[24,2957,2959],{"className":2958},[741],[24,2960,2962,2982],{"className":2961},[84,131],[24,2963,2965,2979],{"className":2964},[88],[24,2966,2968],{"className":2967,"style":751},[92],[24,2969,2970,2973],{"style":2549},[24,2971],{"className":2972,"style":101},[100],[24,2974,2976],{"className":2975},[109,110,111,108],[24,2977,569],{"className":2978,"style":686},[69,632,108],[24,2980,157],{"className":2981},[156],[24,2983,2985],{"className":2984},[88],[24,2986,2988],{"className":2987,"style":773},[92],[24,2989],{}," 就是一个圈；连通分量、孔洞、分叉点——所有这些拓扑性质都被忠实地保留下来。",[11,2992,2993],{},"这一步的意义在于，它第一次用严格的数学语言把「流形长什么样」与「概念本身的结构」联系起来，而不再停留在「看起来像个圆」这种直观描述上。",[224,2995,2996],{"id":2996},"实证验证",[11,2998,2999],{},"再优雅的理论也必须接受数据的检验。作者选取了三个有代表性的案例：",[3001,3002,3003,3011,3014],"ol",{},[1444,3004,3005,3006,3010],{},"颜色：用 OpenAI 的 ",[3007,3008,3009],"code",{},"text-embedding-large-3"," 从英文颜色名称生成 3072 维嵌入，再用 PCA 降到三维用于可视化；",[1444,3012,3013],{},"年份：沿用 Engels 等（2025）的做法，用 SAE 从 GPT-2-small 的第 7 层提取「20 世纪年份」特征；",[1444,3015,3016],{},"日期：用「1 月 1 日」到「12 月 31 日」等提示词生成嵌入，同样使用 OpenAI 的嵌入模型。",[11,3018,3019],{},"结果令人瞩目：颜色嵌入沿一个圆环排列，色相次序（红→紫→蓝→绿→黄→橙→红）与标准色环完全吻合。年份的 token 激活在三维空间中描出一条蜿蜒的曲线，隐约唤起人类对「时间线」的直觉。作者进一步用 k-近邻图估计了流形上的序结构，并与真实年份计算秩相关，得到肯德尔相关系数 0.97、斯皮尔曼相关系数超过 0.99——近乎完美的单调对应，为同胚预测提供了有力支持。",[11,3021,3022],{},"论文里还有一个特别耐人寻味的「陷阱」细节：在 Engels 等人的原始工作中，「星期几」和「月份」的表示投影到前两个主成分上时，呈现出整齐的圆。作者指出，一旦考察第三个主成分就会发现，这个「圆」其实在第三维上持续地扭曲缠绕；被压进二维平面的整齐圆是一种视觉假象（见其图 2）。这一观察是一个告诫：低维投影作为可解释性研究不可或缺的可视化工具，同时也可能误导人。",[224,3024,3025],{"id":3025},"计算表达能力角度",[11,3027,3028],{},"至此一个自然的问题浮现出来：既然「年份」本质上是一维的量，模型为什么不在表征空间中直接把它编码成一条直线段，而要扭曲成一条穿过高维空间的曲线呢？",[11,3030,3031,3032,3060,3061,3137,3138,3266,3267,3295,3296,3357,3358,3517,3518,3562,3563,3624],{},"作者给出的答案既直观又带着数学家的典型气质：表达能力。如果目标仅仅是通过线性投影「读出」 ",[24,3033,3035,3048],{"className":3034,"translate":28},[27],[24,3036,3038],{"className":3037},[32],[34,3039,3040],{"xmlns":36},[38,3041,3042,3046],{},[41,3043,3044],{},[541,3045,1763],{},[48,3047,1763],{"encoding":50},[24,3049,3051],{"className":3050,"ariaHidden":56},[55],[24,3052,3054,3057],{"className":3053},[60],[24,3055],{"className":3056,"style":1026},[64],[24,3058,1763],{"className":3059,"style":1790},[69,632]," 本身（即让恒等函数 ",[24,3062,3064,3093],{"className":3063,"translate":28},[27],[24,3065,3067],{"className":3066},[32],[34,3068,3069],{"xmlns":36},[38,3070,3071,3090],{},[41,3072,3073,3080,3082,3084,3086,3088],{},[41,3074,3075,3078],{},[541,3076,3077],{"mathvariant":543},"i",[541,3079,1327],{"mathvariant":543},[546,3081,549],{"stretchy":548},[541,3083,1763],{},[546,3085,555],{"stretchy":548},[546,3087,558],{},[541,3089,1763],{},[48,3091,3092],{"encoding":50},"\\mathrm{id}(z) = z",[24,3094,3096,3128],{"className":3095,"ariaHidden":56},[55],[24,3097,3099,3102,3110,3113,3116,3119,3122,3125],{"className":3098},[60],[24,3100],{"className":3101,"style":621},[64],[24,3103,3105],{"className":3104},[69],[24,3106,3109],{"className":3107},[69,3108],"mathrm","id",[24,3111,549],{"className":3112},[628],[24,3114,1763],{"className":3115,"style":1790},[69,632],[24,3117,555],{"className":3118},[636],[24,3120],{"className":3121,"style":640},[79],[24,3123,558],{"className":3124},[644],[24,3126],{"className":3127,"style":640},[79],[24,3129,3131,3134],{"className":3130},[60],[24,3132],{"className":3133,"style":1026},[64],[24,3135,1763],{"className":3136,"style":1790},[69,632]," 能通过一次线性运算算出），那么两个正交方向 ",[24,3139,3141,3168],{"className":3140,"translate":28},[27],[24,3142,3144],{"className":3143},[32],[34,3145,3146],{"xmlns":36},[38,3147,3148,3165],{},[41,3149,3150,3157,3159],{},[583,3151,3152,3154],{},[541,3153,600],{},[1967,3155,3156],{},"0",[546,3158,1322],{"separator":56},[583,3160,3161,3163],{},[541,3162,600],{},[1967,3164,1969],{},[48,3166,3167],{"encoding":50},"v_0, v_1",[24,3169,3171],{"className":3170,"ariaHidden":56},[55],[24,3172,3174,3178,3220,3223,3226],{"className":3173},[60],[24,3175],{"className":3176,"style":3177},[64],"height:0.625em;vertical-align:-0.1944em;",[24,3179,3181,3184],{"className":3180},[69],[24,3182,600],{"className":3183,"style":791},[69,632],[24,3185,3187],{"className":3186},[741],[24,3188,3190,3211],{"className":3189},[84,131],[24,3191,3193,3208],{"className":3192},[88],[24,3194,3197],{"className":3195,"style":3196},[92],"height:0.3011em;",[24,3198,3199,3202],{"style":806},[24,3200],{"className":3201,"style":101},[100],[24,3203,3205],{"className":3204},[109,110,111,108],[24,3206,3156],{"className":3207},[69,108],[24,3209,157],{"className":3210},[156],[24,3212,3214],{"className":3213},[88],[24,3215,3218],{"className":3216,"style":3217},[92],"height:0.15em;",[24,3219],{},[24,3221,1322],{"className":3222},[1392],[24,3224],{"className":3225,"style":731},[79],[24,3227,3229,3232],{"className":3228},[69],[24,3230,600],{"className":3231,"style":791},[69,632],[24,3233,3235],{"className":3234},[741],[24,3236,3238,3258],{"className":3237},[84,131],[24,3239,3241,3255],{"className":3240},[88],[24,3242,3244],{"className":3243,"style":3196},[92],[24,3245,3246,3249],{"style":806},[24,3247],{"className":3248,"style":101},[100],[24,3250,3252],{"className":3251},[109,110,111,108],[24,3253,1969],{"className":3254},[69,108],[24,3256,157],{"className":3257},[156],[24,3259,3261],{"className":3260},[88],[24,3262,3264],{"className":3263,"style":3217},[92],[24,3265],{}," 就足够了。但如果还希望让 ",[24,3268,3270,3283],{"className":3269,"translate":28},[27],[24,3271,3273],{"className":3272},[32],[34,3274,3275],{"xmlns":36},[38,3276,3277,3281],{},[41,3278,3279],{},[541,3280,1763],{},[48,3282,1763],{"encoding":50},[24,3284,3286],{"className":3285,"ariaHidden":56},[55],[24,3287,3289,3292],{"className":3288},[60],[24,3290],{"className":3291,"style":1026},[64],[24,3293,1763],{"className":3294,"style":1790},[69,632]," 的高阶多项式（如 ",[24,3297,3299,3318],{"className":3298,"translate":28},[27],[24,3300,3302],{"className":3301},[32],[34,3303,3304],{"xmlns":36},[38,3305,3306,3315],{},[41,3307,3308],{},[1953,3309,3310,3312],{},[541,3311,1763],{},[1967,3313,3314],{},"2",[48,3316,3317],{"encoding":50},"z^2",[24,3319,3321],{"className":3320,"ariaHidden":56},[55],[24,3322,3324,3328],{"className":3323},[60],[24,3325],{"className":3326,"style":3327},[64],"height:0.8141em;",[24,3329,3331,3334],{"className":3330},[69],[24,3332,1763],{"className":3333,"style":1790},[69,632],[24,3335,3337],{"className":3336},[741],[24,3338,3340],{"className":3339},[84],[24,3341,3343],{"className":3342},[88],[24,3344,3346],{"className":3345,"style":3327},[92],[24,3347,3348,3351],{"style":2114},[24,3349],{"className":3350,"style":101},[100],[24,3352,3354],{"className":3353},[109,110,111,108],[24,3355,3314],{"className":3356},[69,108],"）也能被线性读出，就需要更多的正交方向 ",[24,3359,3361,3399],{"className":3360,"translate":28},[27],[24,3362,3364],{"className":3363},[32],[34,3365,3366],{"xmlns":36},[38,3367,3368,3396],{},[41,3369,3370,3376,3378,3381,3383],{},[583,3371,3372,3374],{},[541,3373,600],{},[1967,3375,3156],{},[546,3377,1322],{"separator":56},[546,3379,3380],{},"…",[546,3382,1322],{"separator":56},[583,3384,3385,3387],{},[541,3386,600],{},[41,3388,3389,3391,3394],{},[541,3390,11],{},[546,3392,3393],{},"+",[1967,3395,1969],{},[48,3397,3398],{"encoding":50},"v_0, \\dots, v_{p+1}",[24,3400,3402],{"className":3401,"ariaHidden":56},[55],[24,3403,3405,3409,3449,3452,3455,3459,3462,3465,3468],{"className":3404},[60],[24,3406],{"className":3407,"style":3408},[64],"height:0.7167em;vertical-align:-0.2861em;",[24,3410,3412,3415],{"className":3411},[69],[24,3413,600],{"className":3414,"style":791},[69,632],[24,3416,3418],{"className":3417},[741],[24,3419,3421,3441],{"className":3420},[84,131],[24,3422,3424,3438],{"className":3423},[88],[24,3425,3427],{"className":3426,"style":3196},[92],[24,3428,3429,3432],{"style":806},[24,3430],{"className":3431,"style":101},[100],[24,3433,3435],{"className":3434},[109,110,111,108],[24,3436,3156],{"className":3437},[69,108],[24,3439,157],{"className":3440},[156],[24,3442,3444],{"className":3443},[88],[24,3445,3447],{"className":3446,"style":3217},[92],[24,3448],{},[24,3450,1322],{"className":3451},[1392],[24,3453],{"className":3454,"style":731},[79],[24,3456,3380],{"className":3457},[3458],"minner",[24,3460],{"className":3461,"style":731},[79],[24,3463,1322],{"className":3464},[1392],[24,3466],{"className":3467,"style":731},[79],[24,3469,3471,3474],{"className":3470},[69],[24,3472,600],{"className":3473,"style":791},[69,632],[24,3475,3477],{"className":3476},[741],[24,3478,3480,3509],{"className":3479},[84,131],[24,3481,3483,3506],{"className":3482},[88],[24,3484,3486],{"className":3485,"style":3196},[92],[24,3487,3488,3491],{"style":806},[24,3489],{"className":3490,"style":101},[100],[24,3492,3494],{"className":3493},[109,110,111,108],[24,3495,3497,3500,3503],{"className":3496},[69,108],[24,3498,11],{"className":3499},[69,632,108],[24,3501,3393],{"className":3502},[2131,108],[24,3504,1969],{"className":3505},[69,108],[24,3507,157],{"className":3508},[156],[24,3510,3512],{"className":3511},[88],[24,3513,3515],{"className":3514,"style":773},[92],[24,3516],{},"，相应地 ",[24,3519,3521,3541],{"className":3520,"translate":28},[27],[24,3522,3524],{"className":3523},[32],[34,3525,3526],{"xmlns":36},[38,3527,3528,3538],{},[41,3529,3530,3532,3534,3536],{},[541,3531,1937],{},[546,3533,549],{"stretchy":548},[541,3535,1763],{},[546,3537,555],{"stretchy":548},[48,3539,3540],{"encoding":50},"\\phi(z)",[24,3542,3544],{"className":3543,"ariaHidden":56},[55],[24,3545,3547,3550,3553,3556,3559],{"className":3546},[60],[24,3548],{"className":3549,"style":621},[64],[24,3551,1937],{"className":3552},[69,632],[24,3554,549],{"className":3555},[628],[24,3557,1763],{"className":3558,"style":1790},[69,632],[24,3560,555],{"className":3561},[636]," 描出的路径就必须在一个 ",[24,3564,3566,3588],{"className":3565,"translate":28},[27],[24,3567,3569],{"className":3568},[32],[34,3570,3571],{"xmlns":36},[38,3572,3573,3585],{},[41,3574,3575,3577,3579,3581,3583],{},[546,3576,549],{"stretchy":548},[541,3578,11],{},[546,3580,3393],{},[1967,3582,3314],{},[546,3584,555],{"stretchy":548},[48,3586,3587],{"encoding":50},"(p+2)",[24,3589,3591,3612],{"className":3590,"ariaHidden":56},[55],[24,3592,3594,3597,3600,3603,3606,3609],{"className":3593},[60],[24,3595],{"className":3596,"style":621},[64],[24,3598,549],{"className":3599},[628],[24,3601,11],{"className":3602},[69,632],[24,3604],{"className":3605,"style":964},[79],[24,3607,3393],{"className":3608},[2131],[24,3610],{"className":3611,"style":964},[79],[24,3613,3615,3618,3621],{"className":3614},[60],[24,3616],{"className":3617,"style":621},[64],[24,3619,3314],{"className":3620},[69],[24,3622,555],{"className":3623},[636]," 维子空间中弯曲。",[11,3626,3627],{},"换言之：流形在高维空间中「扭曲」的程度，在某种意义上编码了后续网络层能够通过简单线性投影直接读出多少种关于该特征的不同（非线性）函数。这为「模型为何把简单的一维概念编码成复杂的高维流形」提供了一种功能主义解释——不是因为模型「迫不得已」，而是因为这样做有利于下游计算。",[11,3629,3630,3631,3700],{},"作者随后把这一视角与叠加假说联系起来：由于表征维度远小于潜在特征的数量，模型几乎别无选择，只能让大多数特征在同一个表征空间中稀疏地、近似正交地共享。在这种约束下，",[24,3632,3634,3651],{"className":3633,"translate":28},[27],[24,3635,3637],{"className":3636},[32],[34,3638,3639],{"xmlns":36},[38,3640,3641,3649],{},[41,3642,3643],{},[583,3644,3645,3647],{},[541,3646,1937],{},[541,3648,569],{},[48,3650,2446],{"encoding":50},[24,3652,3654],{"className":3653,"ariaHidden":56},[55],[24,3655,3657,3660],{"className":3656},[60],[24,3658],{"className":3659,"style":1621},[64],[24,3661,3663,3666],{"className":3662},[69],[24,3664,1937],{"className":3665},[69,632],[24,3667,3669],{"className":3668},[741],[24,3670,3672,3692],{"className":3671},[84,131],[24,3673,3675,3689],{"className":3674},[88],[24,3676,3678],{"className":3677,"style":751},[92],[24,3679,3680,3683],{"style":754},[24,3681],{"className":3682,"style":101},[100],[24,3684,3686],{"className":3685},[109,110,111,108],[24,3687,569],{"className":3688,"style":686},[69,632,108],[24,3690,157],{"className":3691},[156],[24,3693,3695],{"className":3694},[88],[24,3696,3698],{"className":3697,"style":773},[92],[24,3699],{},"——即某个特征对应的编码方向——确实包含恒等函数的「线性可读」成分这一事实，也在一定程度上解释了为何在实际中，简单的线性探针往往出奇地擅长从叠加的表征中「钓出」某个特定特征。",[224,3702,3703],{"id":3703},"余弦相似度",[11,3705,3706],{},"如果说前面的章节是脚手架，那么这一节就是论文真正的「硬核」贡献，在我看来也是最值得记住的部分。",[11,3708,3709],{},"作者提出了假说 2（余弦相似度反映距离）：局部来看，表示之间的余弦相似度是特征值之间距离平方的某个递减函数：",[24,3711,3713],{"className":3712,"translate":28},[526],[24,3714,3716,3819],{"className":3715,"translate":28},[27],[24,3717,3719],{"className":3718},[32],[34,3720,3721],{"xmlns":36,"display":535},[38,3722,3723,3816],{},[41,3724,3725,3743,3745,3751,3753,3755,3757,3759,3765,3767,3775,3777,3779,3781,3788,3790,3796,3798,3800,3802,3808,3814],{},[41,3726,3727,3730,3733,3736,3738,3740],{},[541,3728,3729],{"mathvariant":543},"C",[541,3731,3732],{"mathvariant":543},"o",[541,3734,3735],{"mathvariant":543},"s",[541,3737,1957],{"mathvariant":543},[541,3739,3077],{"mathvariant":543},[541,3741,3742],{"mathvariant":543},"m",[546,3744,549],{"stretchy":548},[583,3746,3747,3749],{},[541,3748,1937],{},[541,3750,569],{},[546,3752,549],{"stretchy":548},[541,3754,1763],{},[546,3756,555],{"stretchy":548},[546,3758,1322],{"separator":56},[583,3760,3761,3763],{},[541,3762,1937],{},[541,3764,569],{},[546,3766,549],{"stretchy":548},[1953,3768,3769,3771],{},[541,3770,1763],{},[546,3772,3774],{"mathvariant":543,"lspace":3773,"rspace":3773},"0em","′",[546,3776,555],{"stretchy":548},[546,3778,555],{"stretchy":548},[546,3780,558],{},[583,3782,3783,3786],{},[541,3784,3785],{},"g",[541,3787,569],{},[546,3789,549],{"stretchy":548},[583,3791,3792,3794],{},[541,3793,1327],{},[541,3795,569],{},[546,3797,549],{"stretchy":548},[541,3799,1763],{},[546,3801,1322],{"separator":56},[1953,3803,3804,3806],{},[541,3805,1763],{},[546,3807,3774],{"mathvariant":543,"lspace":3773,"rspace":3773},[1953,3809,3810,3812],{},[546,3811,555],{"stretchy":548},[1967,3813,3314],{},[546,3815,555],{"stretchy":548},[48,3817,3818],{"encoding":50},"\\mathrm{CosSim}(\\phi_f(z), \\phi_f(z')) = g_f(d_f(z, z')^2)",[24,3820,3822,3983],{"className":3821,"ariaHidden":56},[55],[24,3823,3825,3829,3836,3839,3879,3882,3885,3888,3891,3894,3934,3937,3971,3974,3977,3980],{"className":3824},[60],[24,3826],{"className":3827,"style":3828},[64],"height:1.088em;vertical-align:-0.2861em;",[24,3830,3832],{"className":3831},[69],[24,3833,3835],{"className":3834},[69,3108],"CosSim",[24,3837,549],{"className":3838},[628],[24,3840,3842,3845],{"className":3841},[69],[24,3843,1937],{"className":3844},[69,632],[24,3846,3848],{"className":3847},[741],[24,3849,3851,3871],{"className":3850},[84,131],[24,3852,3854,3868],{"className":3853},[88],[24,3855,3857],{"className":3856,"style":751},[92],[24,3858,3859,3862],{"style":754},[24,3860],{"className":3861,"style":101},[100],[24,3863,3865],{"className":3864},[109,110,111,108],[24,3866,569],{"className":3867,"style":686},[69,632,108],[24,3869,157],{"className":3870},[156],[24,3872,3874],{"className":3873},[88],[24,3875,3877],{"className":3876,"style":773},[92],[24,3878],{},[24,3880,549],{"className":3881},[628],[24,3883,1763],{"className":3884,"style":1790},[69,632],[24,3886,555],{"className":3887},[636],[24,3889,1322],{"className":3890},[1392],[24,3892],{"className":3893,"style":731},[79],[24,3895,3897,3900],{"className":3896},[69],[24,3898,1937],{"className":3899},[69,632],[24,3901,3903],{"className":3902},[741],[24,3904,3906,3926],{"className":3905},[84,131],[24,3907,3909,3923],{"className":3908},[88],[24,3910,3912],{"className":3911,"style":751},[92],[24,3913,3914,3917],{"style":754},[24,3915],{"className":3916,"style":101},[100],[24,3918,3920],{"className":3919},[109,110,111,108],[24,3921,569],{"className":3922,"style":686},[69,632,108],[24,3924,157],{"className":3925},[156],[24,3927,3929],{"className":3928},[88],[24,3930,3932],{"className":3931,"style":773},[92],[24,3933],{},[24,3935,549],{"className":3936},[628],[24,3938,3940,3943],{"className":3939},[69],[24,3941,1763],{"className":3942,"style":1790},[69,632],[24,3944,3946],{"className":3945},[741],[24,3947,3949],{"className":3948},[84],[24,3950,3952],{"className":3951},[88],[24,3953,3956],{"className":3954,"style":3955},[92],"height:0.8019em;",[24,3957,3959,3962],{"style":3958},"top:-3.113em;margin-right:0.05em;",[24,3960],{"className":3961,"style":101},[100],[24,3963,3965],{"className":3964},[109,110,111,108],[24,3966,3968],{"className":3967},[69,108],[24,3969,3774],{"className":3970},[69,108],[24,3972,2352],{"className":3973},[636],[24,3975],{"className":3976,"style":640},[79],[24,3978,558],{"className":3979},[644],[24,3981],{"className":3982,"style":640},[79],[24,3984,3986,3990,4030,4033,4073,4076,4079,4082,4085,4117,4147],{"className":3985},[60],[24,3987],{"className":3988,"style":3989},[64],"height:1.1502em;vertical-align:-0.2861em;",[24,3991,3993,3996],{"className":3992},[69],[24,3994,3785],{"className":3995,"style":791},[69,632],[24,3997,3999],{"className":3998},[741],[24,4000,4002,4022],{"className":4001},[84,131],[24,4003,4005,4019],{"className":4004},[88],[24,4006,4008],{"className":4007,"style":751},[92],[24,4009,4010,4013],{"style":806},[24,4011],{"className":4012,"style":101},[100],[24,4014,4016],{"className":4015},[109,110,111,108],[24,4017,569],{"className":4018,"style":686},[69,632,108],[24,4020,157],{"className":4021},[156],[24,4023,4025],{"className":4024},[88],[24,4026,4028],{"className":4027,"style":773},[92],[24,4029],{},[24,4031,549],{"className":4032},[628],[24,4034,4036,4039],{"className":4035},[69],[24,4037,1327],{"className":4038},[69,632],[24,4040,4042],{"className":4041},[741],[24,4043,4045,4065],{"className":4044},[84,131],[24,4046,4048,4062],{"className":4047},[88],[24,4049,4051],{"className":4050,"style":751},[92],[24,4052,4053,4056],{"style":754},[24,4054],{"className":4055,"style":101},[100],[24,4057,4059],{"className":4058},[109,110,111,108],[24,4060,569],{"className":4061,"style":686},[69,632,108],[24,4063,157],{"className":4064},[156],[24,4066,4068],{"className":4067},[88],[24,4069,4071],{"className":4070,"style":773},[92],[24,4072],{},[24,4074,549],{"className":4075},[628],[24,4077,1763],{"className":4078,"style":1790},[69,632],[24,4080,1322],{"className":4081},[1392],[24,4083],{"className":4084,"style":731},[79],[24,4086,4088,4091],{"className":4087},[69],[24,4089,1763],{"className":4090,"style":1790},[69,632],[24,4092,4094],{"className":4093},[741],[24,4095,4097],{"className":4096},[84],[24,4098,4100],{"className":4099},[88],[24,4101,4103],{"className":4102,"style":3955},[92],[24,4104,4105,4108],{"style":3958},[24,4106],{"className":4107,"style":101},[100],[24,4109,4111],{"className":4110},[109,110,111,108],[24,4112,4114],{"className":4113},[69,108],[24,4115,3774],{"className":4116},[69,108],[24,4118,4120,4123],{"className":4119},[636],[24,4121,555],{"className":4122},[636],[24,4124,4126],{"className":4125},[741],[24,4127,4129],{"className":4128},[84],[24,4130,4132],{"className":4131},[88],[24,4133,4136],{"className":4134,"style":4135},[92],"height:0.8641em;",[24,4137,4138,4141],{"style":3958},[24,4139],{"className":4140,"style":101},[100],[24,4142,4144],{"className":4143},[109,110,111,108],[24,4145,3314],{"className":4146},[69,108],[24,4148,555],{"className":4149},[636],[11,4151,4152,4153,4223,4224,4354],{},"对 ",[24,4154,4156,4174],{"className":4155,"translate":28},[27],[24,4157,4159],{"className":4158},[32],[34,4160,4161],{"xmlns":36},[38,4162,4163,4171],{},[41,4164,4165],{},[583,4166,4167,4169],{},[541,4168,3785],{},[541,4170,569],{},[48,4172,4173],{"encoding":50},"g_f",[24,4175,4177],{"className":4176,"ariaHidden":56},[55],[24,4178,4180,4183],{"className":4179},[60],[24,4181],{"className":4182,"style":3408},[64],[24,4184,4186,4189],{"className":4185},[69],[24,4187,3785],{"className":4188,"style":791},[69,632],[24,4190,4192],{"className":4191},[741],[24,4193,4195,4215],{"className":4194},[84,131],[24,4196,4198,4212],{"className":4197},[88],[24,4199,4201],{"className":4200,"style":751},[92],[24,4202,4203,4206],{"style":806},[24,4204],{"className":4205,"style":101},[100],[24,4207,4209],{"className":4208},[109,110,111,108],[24,4210,569],{"className":4211,"style":686},[69,632,108],[24,4213,157],{"className":4214},[156],[24,4216,4218],{"className":4217},[88],[24,4219,4221],{"className":4220,"style":773},[92],[24,4222],{}," 的唯一要求是在 0 附近二次可微且满足 ",[24,4225,4227,4259],{"className":4226,"translate":28},[27],[24,4228,4230],{"className":4229},[32],[34,4231,4232],{"xmlns":36},[38,4233,4234,4256],{},[41,4235,4236,4245,4247,4249,4251,4254],{},[4237,4238,4239,4241,4243],"msubsup",{},[541,4240,3785],{},[541,4242,569],{},[546,4244,3774],{"mathvariant":543,"lspace":3773,"rspace":3773},[546,4246,549],{"stretchy":548},[1967,4248,3156],{},[546,4250,555],{"stretchy":548},[546,4252,4253],{},"\u003C",[1967,4255,3156],{},[48,4257,4258],{"encoding":50},"g_f'(0) \u003C 0",[24,4260,4262,4344],{"className":4261,"ariaHidden":56},[55],[24,4263,4265,4269,4326,4329,4332,4335,4338,4341],{"className":4264},[60],[24,4266],{"className":4267,"style":4268},[64],"height:1.1711em;vertical-align:-0.4192em;",[24,4270,4272,4275],{"className":4271},[69],[24,4273,3785],{"className":4274,"style":791},[69,632],[24,4276,4278],{"className":4277},[741],[24,4279,4281,4317],{"className":4280},[84,131],[24,4282,4284,4314],{"className":4283},[88],[24,4285,4288,4300],{"className":4286,"style":4287},[92],"height:0.7519em;",[24,4289,4291,4294],{"style":4290},"top:-2.4169em;margin-left:-0.0359em;margin-right:0.05em;",[24,4292],{"className":4293,"style":101},[100],[24,4295,4297],{"className":4296},[109,110,111,108],[24,4298,569],{"className":4299,"style":686},[69,632,108],[24,4301,4302,4305],{"style":2114},[24,4303],{"className":4304,"style":101},[100],[24,4306,4308],{"className":4307},[109,110,111,108],[24,4309,4311],{"className":4310},[69,108],[24,4312,3774],{"className":4313},[69,108],[24,4315,157],{"className":4316},[156],[24,4318,4320],{"className":4319},[88],[24,4321,4324],{"className":4322,"style":4323},[92],"height:0.4192em;",[24,4325],{},[24,4327,549],{"className":4328},[628],[24,4330,3156],{"className":4331},[69],[24,4333,555],{"className":4334},[636],[24,4336],{"className":4337,"style":640},[79],[24,4339,4253],{"className":4340},[644],[24,4342],{"className":4343,"style":640},[79],[24,4345,4347,4351],{"className":4346},[60],[24,4348],{"className":4349,"style":4350},[64],"height:0.6444em;",[24,4352,3156],{"className":4353},[69],"；除此之外不再施加任何限制。",[11,4356,4357,4358,4388,4389,4458,4459,4490,4491,4560],{},"在假说 1 与假说 2 同时成立的前提下，作者证明了论文的核心结果定理 1：设 ",[24,4359,4361,4376],{"className":4360,"translate":28},[27],[24,4362,4364],{"className":4363},[32],[34,4365,4366],{"xmlns":36},[38,4367,4368,4373],{},[41,4369,4370],{},[541,4371,4372],{},"η",[48,4374,4375],{"encoding":50},"\\eta",[24,4377,4379],{"className":4378,"ariaHidden":56},[55],[24,4380,4382,4385],{"className":4381},[60],[24,4383],{"className":4384,"style":3177},[64],[24,4386,4372],{"className":4387,"style":791},[69,632]," 是特征空间 ",[24,4390,4392,4409],{"className":4391,"translate":28},[27],[24,4393,4395],{"className":4394},[32],[34,4396,4397],{"xmlns":36},[38,4398,4399,4407],{},[41,4400,4401],{},[583,4402,4403,4405],{},[541,4404,1317],{},[541,4406,569],{},[48,4408,1467],{"encoding":50},[24,4410,4412],{"className":4411,"ariaHidden":56},[55],[24,4413,4415,4418],{"className":4414},[60],[24,4416],{"className":4417,"style":957},[64],[24,4419,4421,4424],{"className":4420},[69],[24,4422,1317],{"className":4423,"style":1353},[69,632],[24,4425,4427],{"className":4426},[741],[24,4428,4430,4450],{"className":4429},[84,131],[24,4431,4433,4447],{"className":4432},[88],[24,4434,4436],{"className":4435,"style":751},[92],[24,4437,4438,4441],{"style":1368},[24,4439],{"className":4440,"style":101},[100],[24,4442,4444],{"className":4443},[109,110,111,108],[24,4445,569],{"className":4446,"style":686},[69,632,108],[24,4448,157],{"className":4449},[156],[24,4451,4453],{"className":4452},[88],[24,4454,4456],{"className":4455,"style":773},[92],[24,4457],{}," 中的一条有限长路径，",[24,4460,4462,4477],{"className":4461,"translate":28},[27],[24,4463,4465],{"className":4464},[32],[34,4466,4467],{"xmlns":36},[38,4468,4469,4474],{},[41,4470,4471],{},[541,4472,4473],{},"γ",[48,4475,4476],{"encoding":50},"\\gamma",[24,4478,4480],{"className":4479,"ariaHidden":56},[55],[24,4481,4483,4486],{"className":4482},[60],[24,4484],{"className":4485,"style":3177},[64],[24,4487,4473],{"className":4488,"style":4489},[69,632],"margin-right:0.0556em;"," 是它在流形 ",[24,4492,4494,4511],{"className":4493,"translate":28},[27],[24,4495,4497],{"className":4496},[32],[34,4498,4499],{"xmlns":36},[38,4500,4501,4509],{},[41,4502,4503],{},[583,4504,4505,4507],{},[541,4506,2513],{},[541,4508,569],{},[48,4510,2518],{"encoding":50},[24,4512,4514],{"className":4513,"ariaHidden":56},[55],[24,4515,4517,4520],{"className":4516},[60],[24,4518],{"className":4519,"style":957},[64],[24,4521,4523,4526],{"className":4522},[69],[24,4524,2513],{"className":4525,"style":2534},[69,632],[24,4527,4529],{"className":4528},[741],[24,4530,4532,4552],{"className":4531},[84,131],[24,4533,4535,4549],{"className":4534},[88],[24,4536,4538],{"className":4537,"style":751},[92],[24,4539,4540,4543],{"style":2549},[24,4541],{"className":4542,"style":101},[100],[24,4544,4546],{"className":4545},[109,110,111,108],[24,4547,569],{"className":4548,"style":686},[69,632,108],[24,4550,157],{"className":4551},[156],[24,4553,4555],{"className":4554},[88],[24,4556,4558],{"className":4557,"style":773},[92],[24,4559],{}," 上的对应路径，则",[24,4562,4564],{"className":4563,"translate":28},[526],[24,4565,4567,4623],{"className":4566,"translate":28},[27],[24,4568,4570],{"className":4569},[32],[34,4571,4572],{"xmlns":36,"display":535},[38,4573,4574,4620],{},[41,4575,4576,4578,4580,4582,4584,4586,4609,4612,4614,4616,4618],{},[541,4577,75],{},[546,4579,549],{"stretchy":548},[541,4581,4473],{},[546,4583,555],{"stretchy":548},[546,4585,558],{},[4587,4588,4589],"msqrt",{},[41,4590,4591,4593,4595,4603,4605,4607],{},[546,4592,1965],{},[1967,4594,3314],{},[4237,4596,4597,4599,4601],{},[541,4598,3785],{},[541,4600,569],{},[546,4602,3774],{"mathvariant":543,"lspace":3773,"rspace":3773},[546,4604,549],{"stretchy":548},[1967,4606,3156],{},[546,4608,555],{"stretchy":548},[546,4610,4611],{},"⋅",[541,4613,75],{},[546,4615,549],{"stretchy":548},[541,4617,4372],{},[546,4619,555],{"stretchy":548},[48,4621,4622],{"encoding":50},"L(\\gamma) = \\sqrt{-2g_f'(0)} \\cdot L(\\eta)",[24,4624,4626,4653,4804],{"className":4625,"ariaHidden":56},[55],[24,4627,4629,4632,4635,4638,4641,4644,4647,4650],{"className":4628},[60],[24,4630],{"className":4631,"style":621},[64],[24,4633,75],{"className":4634},[69,632],[24,4636,549],{"className":4637},[628],[24,4639,4473],{"className":4640,"style":4489},[69,632],[24,4642,555],{"className":4643},[636],[24,4645],{"className":4646,"style":640},[79],[24,4648,558],{"className":4649},[644],[24,4651],{"className":4652,"style":640},[79],[24,4654,4656,4660,4795,4798,4801],{"className":4655},[60],[24,4657],{"className":4658,"style":4659},[64],"height:1.84em;vertical-align:-0.6498em;",[24,4661,4664],{"className":4662},[69,4663],"sqrt",[24,4665,4667,4786],{"className":4666},[84,131],[24,4668,4670,4783],{"className":4669},[88],[24,4671,4674,4760],{"className":4672,"style":4673},[92],"height:1.1902em;",[24,4675,4679,4683],{"className":4676,"style":4678},[4677],"svg-align","top:-3.8em;",[24,4680],{"className":4681,"style":4682},[100],"height:3.8em;",[24,4684,4687,4690,4693,4751,4754,4757],{"className":4685,"style":4686},[69],"padding-left:1em;",[24,4688,1965],{"className":4689},[69],[24,4691,3314],{"className":4692},[69],[24,4694,4696,4699],{"className":4695},[69],[24,4697,3785],{"className":4698,"style":791},[69,632],[24,4700,4702],{"className":4701},[741],[24,4703,4705,4742],{"className":4704},[84,131],[24,4706,4708,4739],{"className":4707},[88],[24,4709,4712,4724],{"className":4710,"style":4711},[92],"height:0.7337em;",[24,4713,4715,4718],{"style":4714},"top:-2.3987em;margin-left:-0.0359em;margin-right:0.05em;",[24,4716],{"className":4717,"style":101},[100],[24,4719,4721],{"className":4720},[109,110,111,108],[24,4722,569],{"className":4723,"style":686},[69,632,108],[24,4725,4727,4730],{"style":4726},"top:-3.0448em;margin-right:0.05em;",[24,4728],{"className":4729,"style":101},[100],[24,4731,4733],{"className":4732},[109,110,111,108],[24,4734,4736],{"className":4735},[69,108],[24,4737,3774],{"className":4738},[69,108],[24,4740,157],{"className":4741},[156],[24,4743,4745],{"className":4744},[88],[24,4746,4749],{"className":4747,"style":4748},[92],"height:0.4374em;",[24,4750],{},[24,4752,549],{"className":4753},[628],[24,4755,3156],{"className":4756},[69],[24,4758,555],{"className":4759},[636],[24,4761,4763,4766],{"style":4762},"top:-3.1502em;",[24,4764],{"className":4765,"style":4682},[100],[24,4767,4771],{"className":4768,"style":4770},[4769],"hide-tail","min-width:1.02em;height:1.88em;",[4772,4773,4779],"svg",{"xmlns":4774,"width":4775,"height":4776,"viewBox":4777,"preserveAspectRatio":4778},"http:\u002F\u002Fwww.w3.org\u002F2000\u002Fsvg","400em","1.88em","0 0 400000 1944","xMinYMin slice",[4780,4781],"path",{"d":4782},"M983 90\nl0 -0\nc4,-6.7,10,-10,18,-10 H400000v40\nH1013.1s-83.4,268,-264.1,840c-180.7,572,-277,876.3,-289,913c-4.7,4.7,-12.7,7,-24,7\ns-12,0,-12,0c-1.3,-3.3,-3.7,-11.7,-7,-25c-35.3,-125.3,-106.7,-373.3,-214,-744\nc-10,12,-21,25,-33,39s-32,39,-32,39c-6,-5.3,-15,-14,-27,-26s25,-30,25,-30\nc26.7,-32.7,52,-63,76,-91s52,-60,52,-60s208,722,208,722\nc56,-175.3,126.3,-397.3,211,-666c84.7,-268.7,153.8,-488.2,207.5,-658.5\nc53.7,-170.3,84.5,-266.8,92.5,-289.5z\nM1001 80h400000v40h-400000z",[24,4784,157],{"className":4785},[156],[24,4787,4789],{"className":4788},[88],[24,4790,4793],{"className":4791,"style":4792},[92],"height:0.6498em;",[24,4794],{},[24,4796],{"className":4797,"style":964},[79],[24,4799,4611],{"className":4800},[2131],[24,4802],{"className":4803,"style":964},[79],[24,4805,4807,4810,4813,4816,4819],{"className":4806},[60],[24,4808],{"className":4809,"style":621},[64],[24,4811,75],{"className":4812},[69,632],[24,4814,549],{"className":4815},[628],[24,4817,4372],{"className":4818,"style":791},[69,632],[24,4820,555],{"className":4821},[636],[11,4823,4824,4825,4894],{},"用大白话说：流形上最短路径（测地线）的长度与特征空间中对应最短路径的长度严格成正比。也就是说，即使不知道 ",[24,4826,4828,4845],{"className":4827,"translate":28},[27],[24,4829,4831],{"className":4830},[32],[34,4832,4833],{"xmlns":36},[38,4834,4835,4843],{},[41,4836,4837],{},[583,4838,4839,4841],{},[541,4840,3785],{},[541,4842,569],{},[48,4844,4173],{"encoding":50},[24,4846,4848],{"className":4847,"ariaHidden":56},[55],[24,4849,4851,4854],{"className":4850},[60],[24,4852],{"className":4853,"style":3408},[64],[24,4855,4857,4860],{"className":4856},[69],[24,4858,3785],{"className":4859,"style":791},[69,632],[24,4861,4863],{"className":4862},[741],[24,4864,4866,4886],{"className":4865},[84,131],[24,4867,4869,4883],{"className":4868},[88],[24,4870,4872],{"className":4871,"style":751},[92],[24,4873,4874,4877],{"style":806},[24,4875],{"className":4876,"style":101},[100],[24,4878,4880],{"className":4879},[109,110,111,108],[24,4881,569],{"className":4882,"style":686},[69,632,108],[24,4884,157],{"className":4885},[156],[24,4887,4889],{"className":4888},[88],[24,4890,4892],{"className":4891,"style":773},[92],[24,4893],{}," 的具体形式，只要局部余弦相似度是距离平方的某个光滑递减函数，「沿着表征流形本身行进」的测地距离，就能精确地——只差一个统一的缩放常数——恢复概念空间中的真实距离。",[11,4896,4897],{},"这一结果回应了 Olah 和 Batson 在 2024 年提出的一个悬而未决的猜想：他们曾写道，「特征流形以超出其拓扑所要求的更复杂方式嵌入，其目的或许正是为了实现某个特定的距离度量，这个想法可能相当深刻而重要。」这篇论文实际上用度量几何的严格语言，把这个直觉性的猜想结晶成一个既可证明也可证伪的定理。",[11,4899,4900],{},"证明本身采用了度量几何中「路径长度」的经典定义（多边形逼近和的上确界），核心技巧是对余弦相似度在零点附近做泰勒展开，随后用一系列仔细的三角不等式论证来界定误差项——一段扎实的解析推理（详见附录）。",[224,4902,4903],{"id":4903},"同胚容易等距困难",[11,4905,4906],{},"理论就位之后，作者回到颜色、年份、日期三个实证案例，检验假说 2 与定理 1。他们设计了两种诊断方法：",[3001,4908,4909,4944],{},[1444,4910,4911,4912,4943],{},"直接方法：以散点图绘制「余弦相似度 vs. 距离平方」，观察是否出现局部递减趋势，并用查特吉相关系数 ",[24,4913,4915,4930],{"className":4914,"translate":28},[27],[24,4916,4918],{"className":4917},[32],[34,4919,4920],{"xmlns":36},[38,4921,4922,4927],{},[41,4923,4924],{},[541,4925,4926],{},"ξ",[48,4928,4929],{"encoding":50},"\\xi",[24,4931,4933],{"className":4932,"ariaHidden":56},[55],[24,4934,4936,4939],{"className":4935},[60],[24,4937],{"className":4938,"style":1214},[64],[24,4940,4926],{"className":4941,"style":4942},[69,632],"margin-right:0.046em;"," 量化整体的函数依赖程度；",[1444,4945,4946],{},"间接方法：检验定理 1 本身——用 k-近邻图估计流形上的测地距离，检查它们是否与特征空间中的测地距离成正比，并用皮尔逊相关系数衡量线性程度。",[11,4948,4949,4950,5013],{},"结果耐人寻味。对于「颜色」和「日期」——两者都天然具有周期结构——一个简单的圆形度量（色相角；一年中的第几天）就能为等距性提供相当有力的支持（日期的皮尔逊相关系数达 0.97）。「年份」却栽了跟头：如果假设年份之间的距离就是 ",[24,4951,4953,4977],{"className":4952,"translate":28},[27],[24,4954,4956],{"className":4955},[32],[34,4957,4958],{"xmlns":36},[38,4959,4960,4974],{},[41,4961,4962,4965,4967,4969,4972],{},[541,4963,4964],{"mathvariant":543},"∣",[541,4966,552],{},[546,4968,1965],{},[541,4970,4971],{},"y",[541,4973,4964],{"mathvariant":543},[48,4975,4976],{"encoding":50},"|x-y|",[24,4978,4980,5001],{"className":4979,"ariaHidden":56},[55],[24,4981,4983,4986,4989,4992,4995,4998],{"className":4982},[60],[24,4984],{"className":4985,"style":621},[64],[24,4987,4964],{"className":4988},[69],[24,4990,552],{"className":4991},[69,632],[24,4993],{"className":4994,"style":964},[79],[24,4996,1965],{"className":4997},[2131],[24,4999],{"className":5000,"style":964},[79],[24,5002,5004,5007,5010],{"className":5003},[60],[24,5005],{"className":5006,"style":621},[64],[24,5008,4971],{"className":5009,"style":791},[69,632],[24,5011,4964],{"className":5012},[69],"（例如 1990 与 2000 相距 10 个单位），经验数据完全不支持这一假说。图 4 显示，越接近「当下」的年份（论文以 GPT-2 的发布年份 2019 为参考点）在流形上被拉得越开——它们的距离被「放大」了。",[11,5015,5016,5017,5096],{},"于是作者做了一个非常优雅的修正：把年份的度量空间换成 ",[24,5018,5020,5050],{"className":5019,"translate":28},[27],[24,5021,5023],{"className":5022},[32],[34,5024,5025],{"xmlns":36},[38,5026,5027,5047],{},[41,5028,5029,5032,5035,5037,5040,5042,5045],{},[541,5030,5031],{},"log",[546,5033,5034],{},"⁡",[546,5036,549],{"stretchy":548},[1967,5038,5039],{},"2019",[546,5041,1965],{},[44,5043,5044],{},"year",[546,5046,555],{"stretchy":548},[48,5048,5049],{"encoding":50},"\\log(2019 - \\text{year})",[24,5051,5053,5081],{"className":5052,"ariaHidden":56},[55],[24,5054,5056,5059,5066,5069,5072,5075,5078],{"className":5055},[60],[24,5057],{"className":5058,"style":621},[64],[24,5060,5062,5063],{"className":5061},[658],"lo",[24,5064,3785],{"style":5065},"margin-right:0.0139em;",[24,5067,549],{"className":5068},[628],[24,5070,5039],{"className":5071},[69],[24,5073],{"className":5074,"style":964},[79],[24,5076,1965],{"className":5077},[2131],[24,5079],{"className":5080,"style":964},[79],[24,5082,5084,5087,5093],{"className":5083},[60],[24,5085],{"className":5086,"style":621},[64],[24,5088,5090],{"className":5089},[69,70],[24,5091,5044],{"className":5092},[69],[24,5094,555],{"className":5095},[636]," 上的欧几里得距离——也就是说，模型编码的或许并不是「日历年」本身，而是对数尺度上的「多久以前」。这个新的度量空间与原年份区间拓扑等价（同胚保持，秩相关仍接近 1），但在这一新度量下检验等距性时，结果从「不支持」翻转为「强烈支持」（查特吉系数 0.84，皮尔逊相关系数 0.99）。",[11,5098,5099],{},"我认为这一节是整篇论文最见洞察力的实证部分：它清晰地表明「同胚」与「等距」是两个完全不同的几何保真度层级。前者只问「形状对不对」；后者问「数值上的距离关系对不对」。一个流形可以与你设想的某个概念空间在拓扑上完全一致，在几何上（距离的具体函数形式）却截然不同。而破解「距离究竟是如何被编码的」这一过程，本身就是对模型如何「理解」时间概念的一次侦察。GPT-2 把「较近的年份」彼此拉得更远，在某种意义上暗示：越接近模型训练截止时间的年份，承载的语义纹理越稠密——新闻和事件挤得更紧——因此被分配了更大的「表征空间预算」。这是一个相当迷人的猜想，尽管论文本身并未深入探讨这一现象背后的因果解释。",[224,5101,5102],{"id":5102},"柏拉图假说",[11,5104,5105],{},"读这篇论文时，我忍不住要把它和另一篇同样是 2024 年发表、在机器学习社区引发大量讨论的立场性论文放在一起比较——Huh、Cheung、Wang 与 Isola 提出的柏拉图式表征假说（PRH）（ICML 2024）。虽然两篇论文的具体研究对象和方法几乎完全不同，但把它们对照起来读，会发现它们是在同一个大问题的两个不同尺度上工作，而且彼此互补——甚至可以视为一条因果链条上的两个相继环节。",[11,5107,5108],{},"PRH 的核心主张是：架构不同、模态不同（视觉、语言）、训练目标不同的深度网络，随着规模和任务多样性的增大，其内部表征会趋向收敛到一个共享的几何结构。作者把这个被越来越多的模型逐步逼近的假想表征空间，类比为柏拉图式的「理想实在」（理念界）——独立于任何具体实体而存在并超越它们：每个具体模型学到的表征，只是对这个「柏拉图式表征」的一个带有噪声和偏差的投影。其实证证据包括：在处理配对的图文数据时，随着模型变大，图像模型与语言模型测量数据点间距离的方式变得越来越相似；在一批架构和训练方式各异的视觉模型中，两两间的表征相似性系统地上升。他们还提出了一个解释性猜想：这种收敛的驱动力在于，模型在学习过程中被迫逼近数据生成过程所共有的真实统计结构（即「实在」本身）——任务越多样，这种收敛压力就越强。",[11,5110,5111],{},"如果说 PRH 关心的是「在不同模型之间，整个表征空间是否正收敛到一个统一的几何」，那么 Modell、Rubin-Delanchy 和 Whiteley 的这篇论文关心的则是更小、更具体的尺度：在单个模型内部，为什么对应于单个特征（颜色、年份、日期）的子流形会呈现特定的几何形状，这个形状与它所对应的「概念」之间精确的数学关系又是什么？",[11,5113,5114],{},"把两篇论文放在一起，可以读出相当连贯的一条逻辑链：",[3001,5116,5117,5120,5123],{},[1444,5118,5119],{},"PRH 提出了一个宏观猜想——不同模型的表征空间正在收敛到一个反映「世界的真实统计结构」的共同几何，而这种收敛的程度可以用「模型间测量数据点距离的一致性」来量化（PRH 论文恰恰使用了核对齐、表征相似性等评估两两距离模式一致性的工具）。但 PRH 本身停留在「存在性」和「收敛趋势」层面，并没有深入一个更基础的问题：在单个模型的表征空间内部，距离本身是如何与概念空间中的「真实」距离相对应的？",[1444,5121,5122],{},"本文《大语言模型中表征流形的起源》回答的恰恰是这个更基础、更微观的问题。定理 1 告诉我们：只要局部余弦相似度与特征距离之间存在某种光滑的函数关系（假说 2），表征流形上的测地距离就能精确地——只差一个比例常数——恢复概念空间中的「真实」距离。这就为 PRH 所声称的「模型以某种一致的方式学会测量数据点之间的距离」提供了一个几何机制层面的解释：模型并不是简单地「记住」距离，而是通过把特征映射到特定形状和曲率的流形上，借助该流形上的测地路径隐式地编码了距离结构。",[1444,5124,5125],{},"反过来，本文观察到的颜色和日期呈现的圆形结构——以及「颜色按标准色环的色相次序排列」这一具体发现——可以看作对 PRH 的微观印证：色相本身是客观存在于物理世界中的周期结构（可见光的连续光谱，连同人类视网膜三类视锥细胞的响应曲线，共同决定了颜色空间）。如果不同的图像模型和不同的语言模型，互不协调、各自独立地都学到「色相是一个圆」，这恰恰是 PRH 所描述的「不同模型收敛到反映世界真实统计结构的共享表征」的一个具体、可分离、可检验的实例。换言之，PRH 提供了「为什么会收敛」的宏观叙事；本文提供了「收敛之后几何长什么样，这个几何又是如何精确编码语义距离」的微观刻画。",[11,5127,5128],{},"当然，两篇工作之间也存在明显的张力，值得坦诚地指出。PRH 的收敛主张在很大程度上依赖跨模型比较——用典型相关分析、核对齐等工具比较不同模型对同一组数据点两两距离模式的测量方式，这属于「表征之间的关系」的统计范畴。而《大语言模型中表征流形的起源》则从头到尾讨论的是单个模型内部、单个特征子空间内的几何结构，几乎不涉及不同模型的表征是否可比较的问题。也就是说，即便「余弦相似度与测地距离之间的定理」对某个特定模型（例如 GPT-2）成立，也不能直接推出另一个架构完全不同的模型会用同样的方式——同样的度量空间、同样的对数尺度——编码「年份」这个特征。事实上，「年份以对数尺度编码，距离随接近参考年份而不断拉伸」这一具体发现，本身就带有相当程度的模型特异性。GPT-2 的参考点天然设定为其自身的发布年份（2019），这表明这种编码方式很可能与这个特定模型训练语料的时间分布强相关，而不是所有语言模型都收敛到的、普遍的「柏拉图式」距离编码。如果在训练截止时间和语料分布都不同的模型上重复同样的实验，对数尺度和参考点会相应地偏移吗？这其实是一个非常值得后续工作直接研究的问题——而且在某种意义上，它构成了对 PRH 收敛主张在「单个特征的特定几何参数」这一更细粒度上进行经验检验的天然实验设计。",[11,5130,5131],{},"依我之见，把这两篇论文放在一起读，比单独读任何一篇都更有启发：PRH 告诉我们「所有人都在朝同一个方向收敛」，而本文则为刻画「收敛到底收敛到什么、这种收敛如何被严格度量与证伪」提供了数学语言。如果未来有人想在具体特征层面真正检验 PRH 是否成立——例如色相的圆形结构、时间的对数编码，是否会在架构和训练数据各不相同的模型中一致地重现——那么《大语言模型中表征流形的起源》中提出的同胚检验、等距检验（查特吉系数、基于 KNN 的测地距离的皮尔逊相关）就构成一套几乎现成的、标准化的工具箱。这或许正是这篇论文除了自身数学结论之外，最具有普遍可复用性的「副产品」。",[224,5133,5134],{"id":5134},"结语",[11,5136,5137],{},"读完这篇论文，我想从几个角度给出我的评价。",[11,5139,5140],{},"首先，这是一次难得的把模糊直觉严格化的努力，其主要价值在于提供了语言和工具，而非一个终极答案。机制可解释性领域目前充斥着「看图说话」式的发现——「哦，这个特征看起来像个圆」「这个看起来像树形结构」——却缺少一种统一的数学语言来组织这些观察。把特征定义为度量空间，并把「同胚」与「等距」这两个概念干净地区分开来，本身就是一件重要的概念工具：它迫使研究者把「我觉得这个流形对应那个概念」这种模糊直觉，翻译成具体、可证伪的假说（「这是不是正确的度量空间？」），并提供具体的统计检验程序（查特吉系数、基于 k-近邻图的测地距离）。这种严谨性目前在领域内相对稀缺。",[11,5142,5143],{},"其次，定理 1 是漂亮的数学，但其「证明力」在相当程度上依赖一个相当强的前提——余弦相似度与距离之间存在光滑的函数关系。这一假说本身并非从模型训练动力学或架构设计推导而来，而是作为一个「合理的猜测」被摆上桌面，再加以经验验证。换言之，这篇论文读起来更像「如果这个假说成立，会有哪些漂亮的推论？」，而非「为什么表征空间必然呈现这种结构」（正是在这个意义上，标题中「起源」一词多少有些一厢情愿——论文并没有从优化目标或梯度下降动力学推导出流形为何「涌现」，它在很大程度上是在「流形已经存在」的前提之下，转而刻画它们应当呈现什么样子）。不过这也并非完全是批评：机制可解释性作为一个整体仍处于「观察现象、构建描述性理论」的阶段，真正从第一性原理推导表征几何的工作少之又少。这篇论文至少把「描述性理论」这一步做得比大多数同类工作更严谨。",[11,5145,5146],{},"第三，年份的对数尺度发现是全篇最令人印象深刻、也最引人深思的实证结果。它暗示了一个值得警惕的方法论陷阱：同胚检验（拓扑对不对？）的门槛非常低，很容易得到「看起来对」的东西，但这远不足以说明我们真正理解了模型的编码方式。如果作者止步于「年份沿一条曲线排列，顺序与真实时间顺序一致，秩相关 0.97，说明模型学会了年份的次序」，这个结论将相当空洞——几乎任何单调映射都能做到这一点。真正携带信息量的是等距检验揭示的距离非均匀拉伸，以及它背后的具体假说：以对数尺度编码「距参考点的时间」。这提醒我们，未来类似的可解释性研究不应止步于「形状看起来像不像？」，而应当像本文一样继续追问「距离度量对不对？」——因为只有如此，才能挖掘出模型真正编码的隐含假设。",[11,5148,5149,5150,5219],{},"第四，从应用的角度看，这篇论文对表征引导（steering）研究具有直接的方法论启示，但距离真正可操作的工具还有一段距离。论文在结尾指出，如果能够学习到 ",[24,5151,5153,5170],{"className":5152,"translate":28},[27],[24,5154,5156],{"className":5155},[32],[34,5157,5158],{"xmlns":36},[38,5159,5160,5168],{},[41,5161,5162],{},[583,5163,5164,5166],{},[541,5165,1937],{},[541,5167,569],{},[48,5169,2446],{"encoding":50},[24,5171,5173],{"className":5172,"ariaHidden":56},[55],[24,5174,5176,5179],{"className":5175},[60],[24,5177],{"className":5178,"style":1621},[64],[24,5180,5182,5185],{"className":5181},[69],[24,5183,1937],{"className":5184},[69,632],[24,5186,5188],{"className":5187},[741],[24,5189,5191,5211],{"className":5190},[84,131],[24,5192,5194,5208],{"className":5193},[88],[24,5195,5197],{"className":5196,"style":751},[92],[24,5198,5199,5202],{"style":754},[24,5200],{"className":5201,"style":101},[100],[24,5203,5205],{"className":5204},[109,110,111,108],[24,5206,569],{"className":5207,"style":686},[69,632,108],[24,5209,157],{"className":5210},[156],[24,5212,5214],{"className":5213},[88],[24,5215,5217],{"className":5216,"style":773},[92],[24,5218],{},"（从特征到流形的映射），原则上就可以进行更精细、尊重概念内在几何结构的表征编辑——例如要让「日期」特征向前平移半年，应该沿流形的测地线行进，而不是在欧几里得空间中简单加一个向量。这是一个有前景的方向，也与作者呼吁的「流形感知的 SAE」相呼应。但目前这仍停留在概念层面：如何稳健地估计一个高维、带噪声的流形，仍然是一个尚未解决的统计问题——作者在局限部分也坦诚地承认了这一点。",[11,5221,5222],{},"第五，一个可能被低估的贡献，是论文对文本嵌入服务的启示。许多从事 RAG、语义检索和推荐系统的从业者，不加思索地把余弦相似度当作「语义接近度」的度量，很少停下来追问这个数值相似度背后究竟潜藏着什么样的几何结构。这篇论文是一个提醒：余弦相似度确实能忠实地反映「沿着概念空间内在几何路径的距离」——但这是一个局部性质，依赖于假说 2 成立，而且不同特征的几何形状截然不同（圆、线段、对数线段）。把它当作跨特征的、普遍的「语义距离」度量，可能会掩盖大量重要的非线性结构（如颜色例子所示，「大距离下的余弦相似度明显偏离等距关系」）。对于调优基于嵌入的检索系统和异常检测管道的工程师来说，这是一个值得牢记的告诫。",[11,5224,5225],{},"这篇论文并不试图解释所有表征几何现象，也没有交付一个可以投入生产部署的工具。它只是做了一份「数学家的活」：对于一个被广泛观察却缺乏精确定义的现象，搭建一个最小但严格的理论框架，并诚实地把它拿到数据面前，检验它的边界在哪里。在一个日益依赖「炼金术式」经验观察的领域，这种回归第一性原理、愿意清晰陈述假设、完整写出证明、并坦诚承认自身局限的工作，本身就很值得被更多人看到。",[11,5227,5228],{},"如果要我说这篇论文给我留下的最深刻印象，大概就是「年份以对数时间距离编码」这一具体发现。它通过一个极小的例子表明，在向量空间看似平淡的几何背后，或许真的存在着某种接近人类认知直觉的东西：越近的事件越清晰、越可精细区分；遥远的事件则在记忆中被压缩。这或许是机制可解释性研究最迷人的地方：不是证明模型是一块会魔法的黑箱，而是把它内部真正在「思考」的东西，一点一点翻译成我们能够读懂的语言。",{"title":195,"searchDepth":196,"depth":196,"links":5230},[5231,5232,5233,5234,5235,5236,5237,5238],{"id":513,"depth":196,"text":513},{"id":1291,"depth":196,"text":1292},{"id":2996,"depth":196,"text":2996},{"id":3025,"depth":196,"text":3025},{"id":3703,"depth":196,"text":3703},{"id":4903,"depth":196,"text":4903},{"id":5102,"depth":196,"text":5102},{"id":5134,"depth":196,"text":5134},"如果你关注了过去两年机制可解释性研究的进展，很可能见过这样一类图：把大语言模型某一层神经元的激活向量投影到二维或三维空间并作图，会发现「月份」「星期几」「年份」「颜色」等概念的表示并不是杂乱无章地散布在空间中，而是沿着一条优美的曲线排列——有时甚至是闭合的圆圈或环面。Chris Olah、Anthropic 的 Josh Batson 以及 Engels 等人，都在包括 并非所有语言模型特征都是一维线性的 在内的研究中展示了这一现象。",{},"\u002Fblog\u002F2026\u002F2026-08-04-the-representations-in-llm",{"title":487,"description":5239},"blog\u002F2026\u002F2026-08-04-the-representations-in-llm","将 LLM 的表征流形形式化：以度量空间定义特征，证明余弦相似度编码了特征之间的测地距离，并在颜色、日期与年份三类数据上实证验证了同胚性与等距性。",[5246,5247],"machine-learning","ai","2026-08-04T00:00:00+08:00","7muZrPVdpjj3Hr2Jo2ta-VJnPzeZr9TL8KvVJunp_nQ",{"left":5251,"top":5251,"width":5252,"height":5252,"rotate":5251,"vFlip":198,"hFlip":198,"body":5253},0,24,"\u003Cpath fill=\"currentColor\" d=\"M16 10c0-2.21-1.79-4-4-4s-4 1.79-4 4s1.79 4 4 4s4-1.79 4-4m-6 0c0-1.1.9-2 2-2s2 .9 2 2s-.9 2-2 2s-2-.9-2-2\"\u002F>\u003Cpath fill=\"currentColor\" d=\"M11.42 21.81c.17.12.38.19.58.19s.41-.06.58-.19c.3-.22 7.45-5.37 7.42-11.82c0-4.41-3.59-8-8-8s-8 3.59-8 8c-.03 6.44 7.12 11.6 7.42 11.82M12 4c3.31 0 6 2.69 6 6c.02 4.44-4.39 8.43-6 9.74c-1.61-1.31-6.02-5.29-6-9.74c0-3.31 2.69-6 6-6\"\u002F>",{"left":5251,"top":5251,"width":5252,"height":5252,"rotate":5251,"vFlip":198,"hFlip":198,"body":5255},"\u003Cpath fill=\"none\" stroke=\"currentColor\" stroke-linecap=\"round\" stroke-linejoin=\"round\" stroke-width=\"2\" d=\"M12 7v5l3 3m6-3a9 9 0 1 1-18 0a9 9 0 0 1 18 0\"\u002F>",{"left":5251,"top":5251,"width":5252,"height":5252,"rotate":5251,"vFlip":198,"hFlip":198,"body":5257},"\u003Cg fill=\"none\" stroke=\"currentColor\" stroke-linecap=\"round\" stroke-linejoin=\"round\" stroke-width=\"2\">\u003Cpath d=\"m4.748 7.645l-.331 3.644c-.05.54-.074.811-.03 1.07a2 2 0 0 0 .238.655c.131.228.325.422.711.808l5.176 5.176c.787.787 1.18 1.18 1.636 1.328c.402.131.835.131 1.237 0c.456-.148.853-.544 1.645-1.336l3.96-3.96c.792-.792 1.187-1.188 1.336-1.644a2 2 0 0 0-.001-1.236c-.148-.457-.543-.853-1.335-1.645l-5.163-5.163c-.39-.39-.584-.584-.813-.716a2 2 0 0 0-.656-.238c-.26-.045-.535-.02-1.084.03l-3.63.33c-.944.086-1.417.129-1.787.335a2 2 0 0 0-.775.775c-.205.368-.248.838-.333 1.773z\"\u002F>\u003Cpath d=\"M9.713 9.713a1 1 0 1 0-1.415-1.414a1 1 0 0 0 1.415 1.414\"\u002F>\u003C\u002Fg>",1789084284139]