[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-embedding-cosine-length-artifact-homogenization-measure":3,"topics-all":38,"news-related-80d2381c-82d3-4200-a481-3b859d3b3e77":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"80d2381c-82d3-4200-a481-3b859d3b3e77","余弦相似度测的同质化,其实是文本长度","复测论文把「AI 让讨论同质化」拉回地面:606 万条 Reddit 评论,用平均两两余弦相似度做度量时,测到的其实是评论长度。截断、安慰剂等六项诊断显示,安慰剂能复现 81% 的不对称,原始 DeBERTa 同质性 99.9% 落在第一个主方向上。修正后,两个阵营在 ChatGPT 之前就已收敛。","去年底有个流传挺广的说法:ChatGPT 之后,美国一个保守派论坛的发言明显变得更趋同,而进步派论坛没有。听起来像「AI 正在磨平人们的表达」的一个硬证据。这篇[复测论文](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21519)给出的答案却是:那个信号量的根本不是观点趋同,是文本长度。\n\n提出这个结论的同一位研究者最近自己重跑了一遍。他的数据是:那个保守派板块里带右翼标记的用户,评论平均长度从 2019 年的 109 词缩到 2022 年的 61 词,然后稳住。评论变短了,而「两两余弦相似度」恰好会随长度下降而上升。\n\n## 一个自己推翻自己的复测\n\n原始研究 [arXiv:2608.13674](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.13674) 用了 606 万条 Reddit 评论,取 r\u002FAskConservatives 和 r\u002FAskALiberal 两个板块 2019 到 2025 年的数据,测量方式是「mean-pooled DeBERTa-v3-base 嵌入之间的平均两两余弦相似度」——注意这个编码器从来没有为句子相似度训练过。\n\n原结论是:2022 年 12 月之后,r\u002FAskConservatives 里右翼用户的组内相似度出现向上断裂,γ = +0.0081,p \u003C 0.001;而 r\u002FAskALiberal 左翼用户「移动得少得多」。\n\n复测从 606 万条里分层抽样 915,225 条(336 个「板块 × 意识形态 × 月份」单元格),换了两个句向量编码器,跑了六项诊断。\n\n## 六个诊断全部指向长度\n\n- **截断实验**:同一批评论从 10 词截到 150 词,平均余弦的上升幅度是原报告断裂值的八倍(0.064 对 0.0081)。\n- **长度安慰剂**:把每条评论换成同长度的随机评论,得到 0.0072,原始估计是 0.0089——长度单独就复现了 81% 的不对称,安慰剂抽样里 58% 达到 p \u003C 0.05。\n- **固定内容模拟**:78% 的效果能被纯长度变化造出来;BGE 下这个比例是 96%(0.0144 对 0.0151);连中心化后那些本就不显著的估计,仍有 66–85% 是长度。\n- **主方向**:原始 DeBERTa 同质性有 99.9% 集中在第一个主方向上——长度信号就坐在那根轴上。\n\n一旦把长度处理掉,不对称在所有编码器下都消失。最粗糙的控制(把平均词数及其与组的交互项加进双重差分)就把 0.0089 (p = 0.021) 压到 0.0024 (p = 0.073);在全部 606 万条上重算,修正后的估计是 0.00 SD (p = 1.00)。\n\n## 换个平台,同样的坑\n\n在 Stack Exchange 上复现核心诊断:8 个站点、2018 年 1 月到 2024 年 3 月、57.5 万条答案。这个指标报出了技术社区「ChatGPT 后更同质」,而安慰剂完整复现了它——注意这里的答案是变长的,方向和 Reddit 相反。一个用来检测「话语收敛」的度量,在 Reddit 上测到评论变短,在 Stack Exchange 上测到答案变长。\n\n## 修正之后还剩下什么\n\n修正之后,两个阵营在 2019 到 2022 年间是收敛的,而且在 ChatGPT 之前就停住了——左右质心距离从 2019 年的 14.5 × 10⁻³ 降到 2022 年的 1.8 × 10⁻³。作者说得很直白:他没有检验为什么两个阵营会收敛,只把这当作对这两个论坛的描述。\n\n有意思的是回复。抽了 19.3 万条直接回复做对照:跨阵营回复比自己人回复更「迎合」对方(右回左 +0.017,左回右 +0.006,右回右 −0.006,p \u003C 0.001),跨阵营回复引用对方的比例是 19.5%,同阵营是 13.6–14.8%。而这种迎合度在 ChatGPT 之后并没有上升。\n\n## 补救措施本身也可能翻车\n\n论文还做了一个反向对照,对象选的是 [Doshi & Hauser 2024](https:\u002F\u002Faclanthology.org\u002F2023.emnlp-main.86\u002F) 那个实验。常见补救是「中心化 + 重新归一化」,但这么处理之后,效应在每个编码器下都反号(BGE −0.39 SD,mpnet −0.23,DeBERTa −0.30)。原因是 AI 辅助写的故事更靠近语料均值,中心化后残差向量只有 0.31–0.61 SD 那么长,再归一化就把「更靠近均值」放大成了噪声。改用不归一化的度量,效应回到 0.29–0.52 SD。\n\n还有一句更狠的:即使长度完全不变、只用固定内容的合成序列,这个时间序列设计也有 10–15% 的运行在 p \u003C 0.05 上拒绝原假设,是名义水平的两到三倍。原因是 84 个月度观测配六个 Newey–West 滞后项,区间偏不保守。\n\n## 为什么这事不止一篇\n\n机制层早有人打过招呼,只是没被这把尺子连起来:[Xiao 等人(EMNLP 2023)](https:\u002F\u002Faclanthology.org\u002F2023.emnlp-main.86\u002F)证明对比学习编码器存在「长度诱导的语义漂移」;[Zhou 等人(ACL 2025)](https:\u002F\u002Faclanthology.org\u002F2025.acl-long.1396\u002F)把它命名为 Length Collapse,给出傅里叶域解释——自注意力相当于低通滤波,文本越长低频成分保留越多,长文本嵌入于是聚成一团。\n\n论文自己也做了一份普查:梳理 31 篇已发表和预印本研究,标出「各向异性编码器」与「时间序列或长度不均衡设计」相交的脆弱位置。\n\n## 该泼的冷水\n\n三点。第一,这是原作者对自己前作的复测,不是独立团队的重复验证;原研究本身在中断时间序列、时间断点回归、趋势校正双重差分上都很扎实——但那些设计吃的是同一条序列,而那条序列量的是长度。第二,论文只覆盖两个平台,伪影的量级甚至符号在其他数据上都会变。第三,议程与声音的拆分只用了单一主方向。\n\n不过作者那个判断我觉得是对的:用长度敏感的度量在 AI 时代找到同质化,很可能只是发现人们写得更短或更长了。那是关于文本体量的发现,关于观点范围还什么都没说。\n\n## 所以呢\n\n真正该带走的不是「AI 没有让人类同质化」,而是:当一个指标变成时间序列,先问它到底在量什么。这篇论文给的诊断顺序挺便宜——先报随机配对基线,再报长度安慰剂,再换几个编码器跑一遍,最后校准检验本身,大多不用重编码。\n\n下次看到「AI 让 XX 变同质了」这类标题,不妨先翻到方法一节,找找它有没有报过文本长度的变化。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.21519","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"4f214978-cac1-4f39-aa4b-f92a0d0934b7","transformer",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"66a88aa5-ee36-47bd-b806-6e655a5d40b2","en","The Homogenization That Was Actually Text Length","Across 6.06 million Reddit comments, mean pairwise cosine tracks length, not convergence: a length-only placebo reproduces 81% of the asymmetry.","Late last year a claim spread widely: after ChatGPT, posts in one conservative American forum became visibly more homogeneous, while a progressive forum showed no comparable change. It read as hard evidence that AI is flattening how people express themselves.\n\nThe researcher who made that claim has now re-run it himself. His answer: the signal was never measuring convergence of ideas at all. It was measuring text length. The right-flaired users in that conservative subreddit saw their average comment shrink from 109 words in 2019 to 61 in 2022, then stabilize. Comments got shorter, and pairwise cosine similarity rises as text gets shorter.\n\n## A Re-measurement That Undoes Its Own Predecessor\n\nThe original study, [arXiv:2608.13674](https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.13674), used 6.06 million Reddit comments from r\u002FAskConservatives and r\u002FAskALiberal spanning 2019 to 2025. The measure was the mean pairwise cosine between mean-pooled DeBERTa-v3-base embeddings, and note that this encoder was never trained for sentence similarity.\n\nThe headline claim: after December 2022, within-group similarity among right-flaired users in r\u002FAskConservatives broke upward against its prior downward trend, with gamma = +0.0081 at p \u003C 0.001, while left-flaired users in r\u002FAskALiberal moved far less.\n\nThe re-measurement draws a stratified sample of 915,225 comments across 336 subreddit-by-ideology-by-month cells, swaps in two sentence-similarity encoders, and runs six diagnostics.\n\n## Every Diagnostic Points at Length\n\n- **Truncation**: truncating the same comments from 10 to 150 words raises mean cosine by eight times the reported break, 0.064 against 0.0081.\n- **Length-only placebo**: swapping every comment for a random one of equal length yields 0.0072 against an observed 0.0089, so length alone reproduces 81% of the asymmetry, and 58% of placebo draws reach p \u003C 0.05.\n- **Fixed-content simulation**: 78% of the effect is manufactured by length variation alone, and under BGE the share is 96%, 0.0144 against 0.0151. Even the centered point estimates that were not significant in the real data are 66-85% length.\n- **Leading direction**: 99.9% of raw DeBERTa homogeneity sits on the first principal direction, which is exactly where the length signal lives.\n\nOnce length is handled, the asymmetry disappears under every encoder. The crudest control, adding mean word count and its interaction with group to the difference-in-differences, cuts the estimate from 0.0089 (p = 0.021) to 0.0024 (p = 0.073). Recomputed on the full 6.06 million comments, the corrected estimate is 0.00 SD (p = 1.00).\n\n## The Same Trap On Another Platform\n\nThe core diagnostics repeat on Stack Exchange: 8 sites, January 2018 to March 2024, 575,000 answers. The measure reports post-ChatGPT homogenization of technical sites, and the placebo fully reproduces it. Note that answers here grew longer, the opposite direction from Reddit. A metric built to detect converging discourse was picking up comments shrinking on Reddit and answers growing on Stack Exchange.\n\n## What Survives The Correction\n\nWith the measure repaired, the two camps converged between 2019 and 2022 and stopped converging before ChatGPT arrived. The Left-Right centroid distance fell from 14.5 x 10^-3 in 2019 to 1.8 x 10^-3 in 2022. The author is blunt about the limit: he has no test of why the camps converged and treats the decline as a description of these two forums.\n\nReplies are the interesting part. Sampling 192,975 direct replies, cross-camp replies accommodate their targets more than same-camp replies do: right-to-left by +0.017, left-to-right by +0.006, and right-to-right by -0.006 (p \u003C 0.001). Cross-camp replies quote their parents 19.5% of the time against 13.6-14.8% within camps. And that accommodation did not rise after ChatGPT.\n\n## Even The Standard Fix Can Flip The Sign\n\nThe paper also runs a reverse control against the [Doshi and Hauser 2024](https:\u002F\u002Faclanthology.org\u002F2023.emnlp-main.86\u002F) experiment. The usual remedy is centering plus renormalization. Applied here, the effect reverses sign under every encoder: -0.39 SD under BGE, -0.23 under mpnet, -0.30 under DeBERTa. AI-assisted stories sit closer to the pooled mean, so after centering their residual vectors are only 0.31-0.61 SD long, and renormalizing them to unit length turns closeness into noise. A measure that does not renormalize recovers the effect at 0.29-0.52 SD.\n\nAnd a harder point: even with no length change at all, running the time-series design on fixed-content synthetic series rejects the null in 10-15% of runs, two to three times the nominal rate, because 84 monthly observations with six Newey-West lags give anti-conservative intervals for a level shift.\n\n## Why This Is Not Just One Paper\n\nThe mechanism was flagged earlier, just never connected to this ruler. [Xiao et al. (EMNLP 2023)](https:\u002F\u002Faclanthology.org\u002F2023.emnlp-main.86\u002F) showed that contrastive-learning encoders suffer length-induced semantic drift. [Zhou et al. (ACL 2025)](https:\u002F\u002Faclanthology.org\u002F2025.acl-long.1396\u002F) named the phenomenon Length Collapse and explained it in the Fourier domain: self-attention acts as a low-pass filter, so longer text keeps more low-frequency components and long-text embeddings cluster together.\n\nThe paper also censuses 31 published and preprinted studies, mapping where anisotropic encoders meet time-series or length-unequal designs.\n\n## The Cold Water\n\nThree caveats. First, this is the original author re-measuring their own earlier work, not an independent replication; the original was careful across interrupted time series, regression discontinuity in time, and trend-corrected difference-in-differences, but every design consumed the same series, and that series measured length. Second, the paper covers two platforms, and the magnitude and even the sign of the artifact will vary with other data and encoders. Third, the agenda-voice split uses a single principal direction.\n\nThe author's own conclusion is the one I find right: a study that finds AI-era homogenization with a length-sensitive measure may be detecting that people write shorter or longer, which is a finding about volume of expression and says nothing yet about the range of ideas.\n\n## So What\n\nThe takeaway is not that AI did not homogenize anything. It is that when a metric becomes a time series, ask what it actually measures first. The diagnostic order here is cheap: report the random-pair baseline, run the length-only placebo, rerun under several encoders, then calibrate the test itself.\n\nSo the next time a headline says AI made some group sound alike, go to the methods section and check whether anyone reported text length.\n","embedding-cosine-length-artifact-homogenization-measure","2026-11-17T11:15:00Z","2026-11-17T11:15:39.616520Z","2026-11-17T11:15:39.616530Z",true,"agent",109,[39,49,58],{"slug":40,"tag_slug":41,"title_zh":42,"title_en":43,"intro_zh":44,"intro_en":45,"id":46,"is_active":35,"created_at":47,"modified_at":48},"image-generation","text-to-image","图像生成：从文生图到逐框编辑的模型竞赛","Image Generation: from text-to-image to frame-level editing","文生图已是红海，战场正转向编辑与可控性：FLUX.2\u002F3 用开源和量化把生图搬进消费级显卡，Qwen-Image、Seedream 等国产模型紧咬不放，GPT Image 2 则把推理能力带进像素。本专题追踪图像生成与编辑模型的发布、架构演进、加速量化与商业化数据，看「好图」如何变成「可用的图」。","Text-to-image is a crowded race, and the frontier has moved to editing and control: FLUX.2\u002F3 push open weights and quantization onto consumer GPUs, Qwen-Image and Seedream close in from China, and GPT Image 2 brings reasoning into pixels. This topic tracks model releases, architecture shifts, acceleration work, and the business numbers behind \"pretty\" becoming \"usable\".","52de7c59-7864-4569-8ae7-3d656d26a1a1","2026-11-20T07:20:23.534506Z","2026-11-20T07:20:23.534527Z",{"slug":50,"tag_slug":50,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":35,"created_at":56,"modified_at":57},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":59,"tag_slug":59,"title_zh":60,"title_en":61,"intro_zh":62,"intro_en":63,"id":64,"is_active":35,"created_at":65,"modified_at":66},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":68},[69,74,79,84,89,94],{"id":70,"title":71,"news_slug":72,"published_at":73},"1c6d5c1b-5693-4633-bb79-8acdd77042e3","同样预算两种噪声:注意力头集中掉线让推理掉12.4点","concentrated-attention-dropout-breaks-llm-reasoning","2026-11-12T01:00:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"f4737ed6-ae99-4869-9d9f-2e3170b4ecd9","Olmo Hybrid 用 token 级 loss 告诉行业：Transformer 和混合架构到底差在哪","olmo-hybrid-token-level-loss-comparison","2026-06-25T22:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"42f5ea83-365c-4a4e-a318-17d7c4c11c15","论文实测:自洽性证书在AIME上24%的题上会误触发","self-consistency-certificates-misfire-correlated-reasoning","2026-11-21T09:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"8e741f04-82a2-46da-9d30-3c31117ca70c","Ambient SFT:让预训练的自己当高温老师","ambient-sft-temperature-matching","2026-11-18T17:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"1dec93aa-e2a2-4b65-941c-6fa04d452175","Lattice横向潜变量:困惑度15.41降到14.40","latent-lattice-language-models-horizontal-latent-perplexity","2026-11-17T09:20:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"4cdb4267-a0c3-423d-8005-504778932ae9","40份技术报告审计:73个主张只有19个有对照","llm-technical-report-ablation-audit","2026-11-17T05:30:00+00:00"]