[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-llm-not-yet-transformed-science":3,"topics-all":39,"news-related-61a2987e-3b76-46c6-9c18-b62c61740733":67},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":25,"news_slug":32,"published_at":33,"created_at":34,"modified_at":35,"is_published":36,"publish_type":37,"image_url":15,"view_count":38},"61a2987e-3b76-46c6-9c18-b62c61740733","250万篇论文实测:LLM还没改变科学","新泽西理工团队分析2015到2025年五个学科的250万篇论文全文：LLM确实改写了词频，796个词显著上冲，delve类词2024年见顶后骤降；但词汇多样性、术语密度、论文结构与合作的长期趋势均未被打断，生物医学主题多样性下降多半是模板套话的测量假象，唯一明确加速的是亚洲计算机论文的作者团队规模。","关于 AI slop 的焦虑已经蔓延进学术界：delve 满天飞、参考文献里混进幻觉条目、编辑担心科学写作正在被模型同质化。新泽西理工学院的四名研究者决定不再空对空争论，直接把问题放到数据上：他们抓取 2015 到 2025 年五个学科的约 250 万篇论文全文，从用词、词汇多样性、术语密度、主题多样性、论文结构到合作网络逐层做 2022 年前后的对照。结论有点反直觉：LLM 改写了科学的词，却还没有改变科学本身。（论文：https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.22975 ）\n\n## 被改写的，只有词\n\n研究者对 14,484 个英文词元逐一拟合带 2022 年断点的分段回归：796 个词在 LLM 出现后显著上冲，199 个显著退潮，71 个先涨后跌。上冲的是 establish、underscore 这类分析腔；退潮的是 good、happen 这类平实口语。最典型的是 delve：2023 年前在论文里几乎绝迹，2024 年冲上峰值，2025 年又骤降，作者把这归因于模型更新和公众对 AI 腔的警觉。信号在计算机学科最为明显，在被公式约束的数学里几乎为零。但镜头拉远看：受影响的词总共一千出头，而语料里的唯一词数以千万计，词频的地壳运动并没有改变科学写作的词汇构成。\n\n## 多样性没塌，是测量在捣乱\n\n那些「科学要完」的指标也经不起推敲。词汇多样性（滑窗型符比 MATTR）在五个学科都没有出现 2022 年的显著斜率变化；术语密度反而在全学科持续上升——按论文的说法，论文显得越来越「科学」，而不是被模型的大白话稀释。生物医学主题多样性在 2022 年后的骤降看着吓人，拆开看，论文之间的相似度大量来自出版模板：权益声明、数据可用性说明、作者贡献表述这些套话。过滤掉 105 个模板短语后，降幅大幅收窄。计算机的主题收拢是真实趋势，但它在 2022 年之前就随学科向深度学习聚拢而开始，2021 年之后甚至略有放缓。\n\n## 结构与合作：老趋势照旧\n\n结构层同样如此。对 437,980 篇计算机论文的章节分析显示：只有 1 到 3 节的短文占比从 17% 降到 4%，8 节以上的长文从 5% 涨到 30%，Related Work 覆盖率从 32% 升到 66%，标准五段式模板的占比从 25% 涨到 55%；引言从平均 2141 词砍到 1236 词，参考文献却从 682 词涨到 1261 词。这些变化都沿着 2022 年之前的老轨迹滑行，LLM 既没加速也没打断。合作网络也延续长期上行。真正的例外只有一个：来自亚洲的计算机论文，作者人数在 2022 年后加速增长；计算机与物理的发文量增长同样被放大。作者把它放进「LLM 降低跨语言协作摩擦」的解释框架，同时强调这仍是老趋势的延续而非新浪潮。\n\n## 交叉印证，与该泼的冷水\n\n这篇论文不是孤证。另一支团队 2025 年 10 月提交的独立研究覆盖 arXiv、bioRxiv、medRxiv、SocArXiv 四个预印本库的 210 万篇文献，结论是 LLM 更像「选择性催化剂」而非全面颠覆者：投稿与修改周期被加速，语言复杂度小幅上升，AI 相关主题不成比例地扩张。两支独立团队、两套数百万级语料，指向同一个方向。（交叉研究：https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.17882 ）\n\n冷水也要泼够：文中所有数字都出自论文自报口径，没有第三方复现；数据止于 2025 年，而作者自己都写明 LLM 渗入研究流程的速度「令人担忧」；论文引用的 Hao 等人 Nature 论文记录的 AI 悖论——个体产出扩张、科学整体焦点收缩——发生在词频统计照不到的层面。词没坏，不等于科学没变。\n\n下次再看到拿 delve 频率暴涨证明科学被 AI 毁掉的说法，先问一句：十年基线在哪？这项研究真正的贡献不是安慰，而是方法——先校准基线，再谈革命有没有发生。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.22975","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,16,19,22],{"id":12,"name":13,"slug":13,"description":14,"color":15},"9112951a-2abb-4214-b63a-385ec7afb2ba","ai-for-science","AI for Science 专题：追踪 AI 在生命科学、化学材料、物理世界模型等科学方向的关键突破",null,{"id":17,"name":18,"slug":18,"description":15,"color":15},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":20,"name":21,"slug":21,"description":15,"color":15},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":23,"name":24,"slug":24,"description":15,"color":15},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[26],{"id":27,"lang":28,"title":29,"summary":30,"content":31},"c846f92c-e73e-40c4-96b2-cb6c7ff6db81","en","2.5M Papers Say LLMs Have Not Yet Transformed Science","A 2.5M-paper, five-discipline audit finds LLMs rewrote word frequencies yet left diversity, terminology, structure and collaboration trends intact.","Panic about AI slop has reached the labs: delve is everywhere, hallucinated references slip into bibliographies, and editors worry that models are homogenizing scientific writing. A four-author team at the New Jersey Institute of Technology chose measurement over debate. They pulled the full texts of roughly 2.5 million papers published from 2015 to 2025 across five disciplines — Computer Science, Math and Physics from arXiv, Health and Life Sciences from PubMed Central — and compared word usage, lexical diversity, terminology density, topic diversity, paper structure and collaboration before and after late 2022. Their verdict: LLMs rewrote the words of science, but have not yet transformed science itself. (Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.22975 )\n\n## Only the words moved\n\nThe team fitted segmented regressions to 14,484 English lemmas with a breakpoint at 2022: 796 lemmas surged after LLMs arrived, 199 declined, and 71 rose then dropped. Surge words such as establish and underscore belong to the analytic register LLMs prefer; decline words such as good and happen are the plain alternatives that model-edited prose avoids. The emblem is delve: nearly absent before 2023, peaking in 2024, falling sharply in 2025 — which the authors attribute to model updates and growing awareness of AI slop. The signal is strongest in Computer Science and close to zero in Math, where formal notation constrains the prose. Zoom out, though: the affected list holds slightly over a thousand words against tens of millions of unique words in the corpus, so the vocabulary of scientific writing kept its composition.\n\n## Diversity did not collapse; the measurement misbehaved\n\nThe scarier indicators do not survive contact either. Lexical diversity, measured by a moving-average type-token ratio, shows no statistically significant slope change in 2022 for any of the five disciplines. The density of scientific terminology keeps rising in all five — papers read as more scientific over time, not more diluted. The post-2022 plunge in biomedical topic diversity looks alarming until you notice where the similarity comes from: publishing metadata, licensing statements and contributor-role boilerplate. Filtering the top 105 boilerplate keyphrases removes most of the decline. The convergence of Computer Science topics is real, but it started well before 2022 as the field coalesced around deep learning, and it has slowed slightly since 2021.\n\n## Structure and collaboration: old trends, same slope\n\nStructure tells the same story. Across 437,980 CS papers, the share with only 1 to 3 sections fell from 17% to 4%, papers with 8 or more sections grew from 5% to 30%, Related Work coverage rose from 32% to 66%, and the standard five-section template grew from 25% to 55% of papers; introductions were cut from 2141 to 1236 words on average while references grew from 682 to 1261 words. Every one of these rides a trend laid down before LLMs. Collaboration kept its long climb in team size after 2022. One exception stands out: the number of coauthors on Computer Science papers from Asia accelerated after 2022, and paper counts in CS and Physics look amplified. The authors place this inside the frame of LLMs lowering cross-lingual friction, while stressing that it extends an old trend rather than starting a new wave.\n\n## Independent corroboration, and cold water\n\nThis is not a lone datapoint. A separate team, whose study of 2.1 million preprints across four repositories (arXiv, bioRxiv, medRxiv, SocArXiv) was submitted in October 2025, concludes that LLMs act less as universal disruptors than as selective catalysts: submission and revision cycles accelerated, linguistic complexity rose modestly, and AI-related topics expanded disproportionately. Two independent teams, two multi-million-document corpora, one direction. (Cross-check: https:\u002F\u002Farxiv.org\u002Fabs\u002F2510.17882 )\n\nThe cold water: every number above is self-reported by the papers, with no third-party replication; the data ends in 2025 while the authors themselves call the speed of LLM penetration concerning; and the AI paradox documented by Hao et al. in Nature — individual output expanding while the collective focus of science contracts — lives on a layer that word statistics cannot referee. The words are fine; that does not settle the science.\n\nSo the next time a soaring delve count is offered as proof that AI ruined science, ask for the baseline. The real contribution of this study is methodological: calibrate a decade of trend first, then argue about whether a revolution happened.","llm-not-yet-transformed-science","2026-11-18T21:40:00Z","2026-11-18T21:08:58.681360Z","2026-11-18T21:08:58.681378Z",true,"agent",79,[40,50,58],{"slug":41,"tag_slug":42,"title_zh":43,"title_en":44,"intro_zh":45,"intro_en":46,"id":47,"is_active":36,"created_at":48,"modified_at":49},"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":13,"tag_slug":13,"title_zh":51,"title_en":52,"intro_zh":53,"intro_en":54,"id":55,"is_active":36,"created_at":56,"modified_at":57},"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":36,"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},"8f0348fd-b4d2-4567-8dae-f2d6f9cf7760","论文实测:算术电路焊进LLM权重,6位数冲上100%","zero-gradient-arithmetic-circuits-llm-weights","2026-11-21T17:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"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":80,"title":81,"news_slug":82,"published_at":83},"34ed8711-148e-4609-972e-2739f3b2d1d5","论文实测:ICLR审稿重叠率12.0%涨到21.3%","iclr-review-critique-overlap-homogenization","2026-11-20T21:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"7726a377-f3e3-4fd1-93eb-6c4b075b685f","论文实测:统一模型自画隐藏状态,代词可视化最高94%","v-lens-unified-model-hidden-state-visualization","2026-11-20T05:30:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"e9ce0c9c-c7f3-48ac-b4e8-bf58587dd451","删掉有害数据就防住了?论文说模型越强补回来越省","filtered-training-data-capability-returns-cheaper-with-scale","2026-11-19T05:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"648ee35f-321e-4cc7-8e88-9a9d3ebae033","小模型每晚自学:一年后从 52 分掉到 5 分","on-device-llm-self-training-judge-failure","2026-11-19T03:30:00+00:00"]