[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-iclr-review-critique-overlap-homogenization":3,"topics-all":36,"news-related-34ed8711-148e-4609-972e-2739f3b2d1d5":64},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":22,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":15,"view_count":35},"34ed8711-148e-4609-972e-2739f3b2d1d5","论文实测:ICLR审稿重叠率12.0%涨到21.3%","单作者论文拆了ICLR八年的299,168条评审:同一篇论文的审稿意见在问题层面重叠率从12.0%涨到21.3%,第三位审稿人带来的新问题从6.12个缩到5.46个,被AI检测器标记的评审对重叠更多,期刊同样在涨。作者结论是评审独立性在流失,但明说这不是LLM使用的因果证据。","同行评审有个默认前提:几个审稿人各自独立读论文,各自踩中不同的坑。1990年有研究观测过,两位审稿人常常连一个共同批评点都没有。这个前提正在动摇:有论文测出,ICLR同一篇论文的几条审稿意见,在问题层面越来越像。\n\n## 重叠率一年涨9.3点\n\n这篇论文(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26286)只有一位作者:Ilya Gusev,供职于Keenable,坐标Amsterdam。他抓了ICLR 2020至2027共85,466篇论文、299,168条评审,其中2027年批次是公开仅八天、还没出决定的评审快照。测法是问题级而非文本级:先用GPT-6 Luna把每条评审拆成一条条批评claim,再把指向同一问题的批评聚成一组,对每对评审算重叠系数。方法本身有校验:对照218条Nature-family评审的人工标注,Luna引用的原文片段落在96.7%的标注条目里;再让Sol 6.1、DeepSeek V4.1 Flash、Kimi K3.1 Pro三个模型独立重跑提取和分组,所有对比方向一致,重叠水平随模型在10.9%到14.8%间浮动。结果:ICLR重叠率2026年12.0%,2027年21.3%,一年涨9.3点;作者强调这两年取的都是讨论阶段评审,口径直接可比。NeurIPS、ICML以及F1000(+5.8点)、PLOS(+2.5点)也在各自最新周期上涨。\n\n更有信息量的是拆账。每条评审的批评条数没降(7.2到7.3),但三条评审去重后的独立问题总数从19.9掉到19.0,比2020年以来任何一年都少;第三位审稿人带来的新增问题从6.12个缩到5.46个;每对评审共享的问题数从0.61翻到1.10。换句话说,字没少写,意见却在收敛。\n\n## 三道反证\n\n作者用MELD检测器(阈值取110,080条2023年前评分的99.9分位)标记疑似有AI参与的评审,发现同一篇论文里,双标记评审对的重叠比双未标记高:ICLR 2024-2025高3.0点,2026-2027高4.5点,五个期刊源合计高5.4点。反证做了三层:把600组人类评审只做语言润色,重叠只升0.5点;按提纲重写,也只升0.9点——都远够不着3.0到4.5点的差距。在ChatGPT之前的ICLR数据上强行标记最高45%,差距落在-3.3到+0.2,说明检测器高分本身造不出差距。换GigaCheck检测器重跑,结论方向不变。\n\n## 作者自己泼的冷水\n\n照论文原话转述的几条限定:涨的9.3点里,双未标记评审对只能解释3.7点,其余5.6点跟着双标记评审走,而论文明说这是记账,不是LLM使用的因果效应;被标记不等于机器代写,润色过的评审也有39.5%会被标;未被标记也不等于纯人写;检测器标签不能拿来针对任何个体审稿人;NeurIPS和多数期刊只公开录用论文的评审。论文留了两个开放问题:多少涨幅来自不留文本痕迹的LLM使用,以及2027年的水平到讨论阶段结束后是否还成立。顺带一提,这篇论文自己的代码和文本由Claude Fable 5.2和GPT-6 Astra完成,作者的角色是监督和编辑。\n\n## 所以呢\n\n机制层这不是孤例:创意写作领域早有同款发现,Anderson等人(https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.01536)测过LLM辅助让不同人的点子趋同,Doshi与Hauser在Science Advances报过个体创意上升、集体多样性下降。论文还引用了Ghosh等人的研究(https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17811):AI评审超过一半的弱点来自附录,ICLR抽样里这类占比从11.3%涨到27.6%。这一现象其实早有铺垫:Liang等人2024年就在ICML上量化过ChatGPT对会议评审文本的渗透,产业侧Pangram也判过21%的ICLR评审疑似AI生成,这篇新工作只是把粒度推进到了问题级。真正扎心的是政策盲区:ICML 2026的随机政策试验只盯长度和分数,评审变长5.5-7%,分数几乎不动,没人测意见多样性。同行评审的价值恰恰建立在几个读者发现不同问题上;当第三位审稿人的新增问题缩水,多一个人审稿的边际价值就在缩水。对投稿人的启示很直接:别默认三条审稿意见覆盖了三份盲区,它们可能正在收敛成同一份。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26286","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,16,19],{"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},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":20,"name":21,"slug":21,"description":15,"color":15},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[23],{"id":24,"lang":25,"title":26,"summary":27,"content":28},"52c6b90a-3989-40a2-8a5c-1682657a7dc5","en","ICLR review critique overlap rises from 12.0% to 21.3%","Study of 299,168 ICLR reviews: critique overlap per paper rose from 12.0% to 21.3%; third reviews add fewer new problems; flagged pairs overlap more.","Peer review rests on a default assumption: several reviewers read a paper independently and step on different problems. A 1990 study observed that two reviewers often shared no critical point at all. That assumption is now shaking: a paper measures that the reviews of one ICLR paper increasingly repeat each other at the level of problems.\n\n## Overlap up 9.3 points in one year\n\nThe paper (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26286) has a single author, Ilya Gusev of Keenable, Amsterdam. It covers 85,466 ICLR papers and 299,168 reviews from 2020 to 2027; the 2027 batch is a snapshot taken eight days after the reviews became public and before decisions. The measurement is claim-level, not whole-text similarity: GPT-6 Luna first splits each review into critique claims, claims pointing at the same problem are grouped, and an overlap coefficient is computed for each pair. The pipeline is itself validated: against human labels on 218 Nature-family reviews, Luna's quotes fall inside 96.7% of the labeled items; Sol 6.1, DeepSeek V4.1 Flash and Kimi K3.1 Pro independently redo extraction and grouping, every contrast keeps its sign, and overlap levels float between 10.9% and 14.8% depending on the model. ICLR overlap rose from 12.0% in 2026 to 21.3% in 2027, up 9.3 points in one year; the author stresses both years hold discussion-phase reviews, so the comparison is directly comparable. NeurIPS, ICML, F1000 (+5.8 points) and PLOS (+2.5 points) also rose in their latest cycles.\n\nThe deeper accounting is more telling. Claims per review did not fall (7.2 to 7.3), but distinct problems across three reviews dropped from 19.9 to 19.0, fewer than any year since 2020; the new problems a third review adds shrank from 6.12 to 5.46; shared problems per pair jumped from 0.61 to 1.10. In short, reviewers write as much, but their critiques converge.\n\n## Three counter-checks\n\nUsing the MELD detector (threshold at the 99.9th percentile of 110,080 pre-2023 scores), the author flags reviews likely touched by AI. Within one paper, pairs of two flagged reviews overlap more than pairs of two unflagged ones: by 3.0 points at ICLR 2024-2025, by 4.5 points in 2026-2027, and by 5.4 points pooled over five journal sources. Three counter-checks follow. Polishing the language of 600 pairs of human reviews raises overlap by only 0.5 points; rewriting each from an outline of its points, by 0.9 points - both far below the 3.0-to-4.5-point gap. Forcing a top-45% flag rate on pre-ChatGPT ICLR data yields gaps of -3.3 to +0.2, so a high detector score alone does not produce the gap. Re-running with the GigaCheck detector keeps the direction.\n\n## The cold water the author pours\n\nLimitations stated in the paper itself: of the 9.3-point rise, unflagged pairs account for only 3.7 points while the other 5.6 points go with flagged reviews, and the paper says this split is an accounting, not a causal effect of LLM use. A flag does not mean machine-written - 39.5% of polished reviews get flagged; an unflagged review is not thereby human-written. Detector labels must not support action against an individual reviewer. NeurIPS and most journals publish reviews of accepted papers only. Two questions stay open: how much of the rise comes from LLM use that leaves no trace in the text, and whether the 2027 level holds after the discussion phase. For the record, the code and text of this very paper were written by Claude Fable 5.2 and GPT-6 Astra, with the authors as supervisors and editors.\n\n## So what\n\nThe mechanism is not an isolated case: Anderson et al. (https:\u002F\u002Farxiv.org\u002Fabs\u002F2402.01536) found LLM assistance homogenizes ideas across writers, and Doshi and Hauser reported in Science Advances that individual creativity rises while collective diversity falls. The paper also cites Ghosh et al. (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.17811): AI reviewers draw more than half of their weaknesses from the appendix, and in sampled ICLR reviews that share rose from 11.3% to 27.6%. Nor is this new: Liang et al. quantified the spread of ChatGPT-modified conference review text at ICML back in 2024, and Pangram judged 21% of ICLR reviews likely AI-generated; this work simply pushes the granularity down to problems. The uncomfortable policy blind spot: the ICML 2026 randomized policy trial tracked length and scores - reviews got 5.5-7% longer with scores almost unchanged - but nobody measured diversity of critiques. The value of peer review rests exactly on several readers finding different problems; as the third reviewer adds fewer new ones, the marginal value of an extra reviewer shrinks. For authors the lesson is blunt: do not assume three reviews cover three blind spots - they may be converging into one.","iclr-review-critique-overlap-homogenization","2026-11-20T21:30:00Z","2026-11-20T21:11:35.555493Z","2026-11-20T21:11:35.555501Z",true,"agent",30,[37,47,55],{"slug":38,"tag_slug":39,"title_zh":40,"title_en":41,"intro_zh":42,"intro_en":43,"id":44,"is_active":33,"created_at":45,"modified_at":46},"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":48,"title_en":49,"intro_zh":50,"intro_en":51,"id":52,"is_active":33,"created_at":53,"modified_at":54},"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":56,"tag_slug":56,"title_zh":57,"title_en":58,"intro_zh":59,"intro_en":60,"id":61,"is_active":33,"created_at":62,"modified_at":63},"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":65},[66,71,76,81,86,91],{"id":67,"title":68,"news_slug":69,"published_at":70},"61a2987e-3b76-46c6-9c18-b62c61740733","250万篇论文实测:LLM还没改变科学","llm-not-yet-transformed-science","2026-11-18T21:40:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"1b3b7024-b8c2-43f5-93bb-3d838ed87f17","GPT-Rosalind：OpenAI 推出首款生命科学专用推理模型","gpt-rosalind-openai-life-science-bixbench","2026-04-29T07:01:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"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":82,"title":83,"news_slug":84,"published_at":85},"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":87,"title":88,"news_slug":89,"published_at":90},"53dd30bb-ba19-45b6-a2de-5011709608b6","论文实测:一段文本认出39%审稿人,靠的是曝光量","frontier-models-name-authors-by-exposure","2026-11-21T01:40:00+00:00",{"id":92,"title":93,"news_slug":94,"published_at":95},"f9cabba1-6f17-4aef-8a62-2bb865468313","论文实测:后训练改变模型读法,知识没变","post-training-changes-readout-not-representations","2026-11-20T17:10:34+00:00"]