[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-frontier-models-name-authors-by-exposure":3,"topics-all":38,"news-related-53dd30bb-ba19-45b6-a2de-5011709608b6":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},"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","论文实测:一段文本加一句零样本提示,模型能认出 39% 公开同行评审的作者,在 90 篇 Nature 系列评审上认出 78%。真正的变量不是文风而是曝光量:Reddit 超过 500 帖的账号被认出率 68.9%,模型自评熟悉度的曲线下面积 0.89,高于用发帖活跃度预测的 0.78。","2026 年 5 月，文学杂志 Granta 收到一篇小说投稿，问 Claude 一句「这是不是 AI 写的」，然后被同行嘲笑太天真——一篇文章的作者身份，不该靠一次模型问答来定。半年后，一篇论文给这件事补了个更冷的注脚：Granta 并不天真，只是用早了几个月的模型。\n\n论文的设定朴素得有点冒犯：给一段文本，加一句零样本提示，不给候选名单，不联网，不要工具，不附写作样本，然后问三个层层收紧的问题——是人写的还是机器写的？如果是机器，哪家实验室写的？如果是人，具体是谁？Claude Fable 5.2 在公开评审里认出 39% 的作者，在 90 篇 Nature 系列评审上认出 78%；Opus 5.5 和 Gemini 4 Argon 对 Anthropic、OpenAI、Google 三家生成的文本有 98% 能报出正确实验室，换成另外八家实验室则最高只有 35%。\n\n论文全文见 arXiv 2611.26216 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26216)。\n\n## 决定认不认得出的，是曝光量\n\n真正反直觉的部分不在准确率，在梯度的形状。论文用 Reddit 活跃度切了五档：完全没有活动历史的账号，认出率 2.1%；发过 1 到 50 条的 12.8%；51 到 200 条 33.1%；201 到 500 条 53.6%；超过 500 条的 68.9%。另一条线用 Goodreads 累计评分当知名度代理——没有作者主页的故事 4.2%，评分过 10k 的作者 100%。\n\n这说明模型不是在「分析文风」，而是在回忆自己读过什么。论文自己的措辞很直接：模型只依赖参数化知识，对它完全不认识的作者，成功率不到 2%。换句话说，一个人能不能被认出来，主要不取决于他怎么写，而取决于他写过多少、公开在哪。\n\n## 模型可以先告诉你它有多了解你\n\n更实用的一层是自评。让模型只报一个 0 到 3 的「我有多了解这个人」，这个自评分预测归属准确率的曲线下面积是 0.89，高于用发帖活跃度预测的 0.78。模型报置信度 25% 以上的回答里有 87% 是对的；同时它系统性低估自己——公开数据上平均自报置信度 17%，实际准确率 27%。按置信度筛一遍，高置信子集几乎可以直接当结论用。\n\n防御这一侧则要分开看。用 AI 把文本改写一遍确实有用：私人评审上 Opus 5.5 的归属准确率从 41.2% 掉到 16.1% 和 15.1%，但论文点明改写只挡表层线索，内容本身的线索还在，对模型本来就熟的人效果更差。指望模型拒答基本落空——拒答主要出现在不足 2% 的文档上，Anthropic 一侧拦下 Fable 5.2 的 20 篇公开评审、3 条 Reddit 评论和 1 篇论文，Opus 5.5 是 28 篇和 2 篇。\n\n还有一个更值得警惕的对称性：认出和模仿用的是同一份记忆。论文让 Opus 5.5 模仿 569 个账号的写作风格再让它辨认，对 166 个「自评熟悉且能从真实评论认出」的账号，它有 74 个（45%）仍被判成目标；而对 219 个熟悉却认不出的账号，只有 9 个（4%）。\n\n## 三条冷下来的注\n\n第一，私有评审那 398 篇来自论文作者自己：四位共同作者加上 UMD CLIP Lab 的两名博士生，一共 6 个写作者，投过 26 个双盲的 ML 与 NLP 会场。56% 这个数只覆盖这 6 个人，论文自己也在正文里说明，规模更大的匿名评审样本要等后续更新。\n\n第二，两个口径别混。摘要里的 39% 对应附录表格中「未被判为 AI」的 789 篇（39.3%）；若把全部 1,559 篇公开评审算进来是 21.4%。这正是论文自定的口径——公开评审默认只在未被判为 AI 的文本上报告，图注也写明打印的数字是非 AI 文本的准确率。剔掉 AI 判定文本后，公开评审的准确率对每个模型相对提升 75% 到 87%，Opus 5.5 从 15% 升到 28%。两个数都是论文自己的，只是分母不同。\n\n第三，比对对象与作者团队有重叠：论文拿来对照的 Pangram 4.1 和 IdeaLens，作者与本文有交集，Bradley Emi 在 Pangram Labs，Rishanth Rajendhran 是本文共同作者。论文做了自审并报告「没有检测器在自己实验室的文本上更准」，这对论文是加分项，但读者仍该知道对照面长什么样。\n\n## 换个域再看一遍\n\n机制层的结论有独立团队在别的域上复现过。CMU Africa 的 Gloire 与 Mitra 测的是双盲评审 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.05157)：给定 5 个同领域候选作者，Qwen2.5-72B 的 top-1 准确率 42.34%，人类研究者只有 11%，平均倒数排名 0.623 对 0.36，在 50% 置信阈值下候选集收到 1.95 个、仍有 61.6% 的情况保留真作者。Lermen 等人的大规模线上去匿名化工作 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.16800) 则做到了 68% 召回、90% 精度，但那是靠联网加候选池堆出来的——本文明确把「不需要候选名单和联网」当作与前作的差别。\n\n反方向也有提醒。Barkhordar 与 Thapa 在代码归属上发现 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.30048)，15 组模型与基准的组合里平衡准确率只有 49% 到 58%，而两两对比的正确率与「评审者自己的方案有多长」的相关系数高达 0.93；剥掉文档字符串、注释、类型标注和局部变量名之后，12 个重测结果里有 10 个掉回随机水平。零样本归属里有多少是表面特征，值得同样拿来追问人类作者这一侧。防御侧，Fisher 等人早就在做「改写」这条路 (https:\u002F\u002Faclanthology.org\u002F2024.naacl-long.87\u002F)，但那类方法挡不住更近的模型。\n\n## 所以呢\n\nGloire 与 Mitra 那句结论值得抄下来：双盲评审应当被理解成概率性保障，而不是匿名性的保证。放到更日常的场景，这篇论文给的判断标准其实很朴素——决定模型能不能认出你的，不是你的文风，是你留在公开语料里的字数；所以最省事的防御不是改写句子，是少留文本。\n\n代价是，你没法知道模型到底读过你多少。你只能自己问它——而它答得相当诚实。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26216","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":22,"name":23,"slug":23,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"2488f9bb-7434-4eae-81eb-efeab5f5911c","en","Paper test: 39% of peer review authors named from one text","Paper test: models name 39% of peer review authors from one text, 78% on a Nature subset. Exposure decides it: 68.9% for active Reddit users.","In May 2026 the literary magazine Granta asked Claude one question about a submitted story: is this AI-written? Reviewers mocked the magazine for it. A paper that landed six months later offers a colder footnote on that episode: Granta was not naive, it just had a model a few months early.\n\nThe setup is almost insultingly simple. Give the model a piece of text and one zero-shot prompt. No candidate list, no web access, no tools, no writing samples. Then ask three progressively tighter questions: was this written by a human or a machine? If a machine, which lab? If a human, which person? Claude Fable 5.2 names the author of 39% of public peer reviews and 78% of a 90-review Nature-family subset. Opus 5.5 and Gemini 4 Argon name the correct lab for 98% of texts from Anthropic, OpenAI and Google, and at most 35% for the other eight labs.\n\nFull paper: arXiv 2611.26216 (https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.26216).\n\n## The deciding variable is exposure, not style\n\nThe counterintuitive part is not the accuracy, it is the shape of the gradient. The paper splits Reddit accounts into five activity buckets: accounts with no activity history are identified 2.1% of the time, those with 1 to 50 posts 12.8%, 51 to 200 33.1%, 201 to 500 53.6%, and over 500 posts 68.9%. A second line uses Goodreads ratings as a fame proxy: stories by authors with no Goodreads profile come out at 4.2%, and authors with ratings over 10k at 100%.\n\nThat is not the model analysing prose style. That is the model recalling what it has read. The paper's own wording is blunt: the model relies solely on its parametric knowledge, and on authors it says it does not know at all, accuracy is under 2%. Whether you can be identified depends far less on how you write than on how much you have written, and where it is public.\n\n## The model tells you how well it knows you first\n\nThe more useful layer is self-report. Ask the model to rate on a 0-to-3 scale how well it knows a given person, and that self-rating predicts attribution accuracy with an area under the curve of 0.89, against 0.78 for using posting activity instead. When the model reports confidence of 25% or above, 87% of those answers are right. It also systematically understates itself: on public data its mean reported confidence is 17% while accuracy is 27%. Filter on confidence and the surviving high-confidence subset is close to usable as-is.\n\nThe defensive side needs separating. Rewriting a text with AI does work: on private peer reviews Opus 5.5's attribution accuracy drops from 41.2% to 16.1% and 15.1%. But as the paper notes, rewriting only blunts surface cues, content-related cues persist, and authors the model already knows well fare worse. Expecting refusals mostly fails: refusals land on under 2% of all documents, with Anthropic blocking Fable 5.2 on 20 public peer reviews, 3 Reddit comments and 1 paper, and Opus 5.5 on 28 peer reviews and 2 papers.\n\nThere is a symmetry worth worrying about more: recognising and imitating draw on the same memory. The paper had Opus 5.5 imitate the writing style of 569 accounts and then try to identify them. For 166 accounts it both rates familiar and identifies from real comments, it still names the target from at least one imitation in 74 cases (45%); for 219 familiar accounts it fails to identify, only 9 (4%).\n\n## Three cold notes\n\nFirst, the 398 private peer reviews came from the paper's own authors: four coauthors plus two PhD students in the UMD CLIP Lab, 6 writers in total, whose reviews went to 26 double-blind ML and NLP venues. That 56% covers only those 6 people, and the paper says in the text that a much larger anonymous review set is still to come.\n\nSecond, do not mix the two denominators. The 39% in the abstract corresponds to the 789 reviews not labelled AI in the appendix table (39.3%); across all 1,559 public reviews the number is 21.4%. That is the paper's own reporting convention, since public reviews and arXiv papers are scored by default only on texts not labelled AI, and the figure caption states the printed numbers are accuracy on non-AI labelled texts. Removing AI-labelled texts raises public-review accuracy by 75% to 87% relative for every model, with Opus 5.5 going from 15% to 28%. Both numbers are the paper's own; they just have different denominators.\n\nThird, the comparison targets overlap with the author team: Pangram 4.1 and IdeaLens share authors with this paper, Bradley Emi is at Pangram Labs and Rishanth Rajendhran is a co-author. The paper audits this and reports that no detector does better on its own lab's texts, which is a point in its favour, but readers should still know what the comparison surface looks like.\n\n## The same mechanism, another domain\n\nIndependent teams reproduce the mechanism-level claim elsewhere. Gloire and Mitra at CMU Africa test double-blind review (https:\u002F\u002Farxiv.org\u002Fabs\u002F2608.05157): given 5 domain-expert candidates, Qwen2.5-72B reaches 42.34% top-1 accuracy where human researchers manage 11%, with a mean reciprocal rank of 0.623 against 0.36, and at a 50% confidence threshold the suspect set shrinks to 1.95 candidates while still retaining the true author 61.6% of the time. Lermen and colleagues' large-scale online deanonymization work (https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.16800) reaches 68% recall at 90% precision, but that is built on web access plus candidate pools, and this paper explicitly frames its difference as needing neither candidate lists nor the web.\n\nThere is a counterweight too. On code attribution, Barkhordar and Thapa find (https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.30048) that balanced accuracy across 15 model-benchmark combinations is only 49% to 58%, and that pairwise accuracy correlates at 0.93 with how much longer the evaluator's own solution is; after stripping docstrings, comments, type hints and local names, 10 of 12 re-tested results fall back to chance. How much of zero-shot attribution is surface signal deserves the same question on the human-author side. On the defensive side, Fisher and colleagues have worked on rewriting as a defence for a while (https:\u002F\u002Faclanthology.org\u002F2024.naacl-long.87\u002F), though those methods do not hold up against more recent models.\n\n## So what\n\nGloire and Mitra's conclusion is worth copying: double-blind review should be understood as a probabilistic safeguard rather than a guarantee of anonymity. For ordinary scenarios this paper's criterion is plain: what decides whether a model can identify you is not your style but how many words you left in the public corpus. So the cheapest defence is not rewriting sentences, it is writing less.\n\nThe cost is that you cannot know how much of you the model has read. You can only ask it, and it answers fairly honestly.","frontier-models-name-authors-by-exposure","2026-11-21T01:40:00Z","2026-11-21T01:24:31.068327Z","2026-11-21T01:24:31.068339Z",true,"agent",28,[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},"251f5c18-2931-4f22-ab50-e19b7d5f58a9","AUROC 0.999 也可能零预警:监控器该怎么评","prewarn-token-level-monitor-early-warning","2026-11-15T05:20:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"243b1b4b-d84b-47a6-a7b1-e252d9f6f3fb","18.9%的英文旧闻被检出近重复:AI评委把通稿当三方佐证","copies-not-confirmations-news-duplication-llm-judges","2026-11-06T09:22:29+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"cc5edf10-22b4-493e-a426-e376ce244418","训练数据暗藏副作用:人类猜不准,模型猜得准","training-data-surprising-side-effects-dsse","2026-11-05T11:16:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"a2ec5815-edbe-4467-a6ed-e3251b09023a","多轮越狱防御新解:把浅层对齐缺陷改成防御支点","shallowdefend-multi-turn-jailbreak-defense","2026-11-01T11:20:00+00:00",{"id":90,"title":91,"news_slug":92,"published_at":93},"0da59714-49b8-4ea4-a74e-fbac3e8c532f","实测18个主流模型:财务问答平均57%答错,难题88%","saturn-ai-financial-advice-error-rate","2026-09-21T17:30:00+00:00",{"id":95,"title":96,"news_slug":97,"published_at":98},"b05de01b-89ca-499b-b130-e55162e651f5","SCOPE：让大模型学会选择性信任，而不是把上下文一概拒绝","scope-selective-trust-context-dpo","2026-08-06T17:59:58+00:00"]