[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-copies-not-confirmations-news-duplication-llm-judges":3,"topics-all":38,"news-related-243b1b4b-d84b-47a6-a7b1-e252d9f6f3fb":57},{"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},"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","论文用MinHash扫描20万篇CC-News英文报道,18.9%能连上别家媒体的近重复稿,人工标注显示最多的是通稿与同集团共享。把一条假消息伪装成三家媒体报道,phi4接受谎言的比例从0.21跳到0.90,加提示说同源也压不住,只有Gemini 2.5 Pro照来源标注走。解法是先按重复链去重,每链只留一份。","一条通稿能同时出现在几十家门户上。人类读者知道这是同一条新闻,读新闻流的模型却看到几十个独立信源。arXiv 论文 2611.09203 量化了这个落差。\n\n## 复制不是佐证:近两成旧闻是别家的副本\n\n论文从 CC-News(2017—2019)抽了 20 万篇英文报道,涉及 6877 家媒体,用 MinHash 加局部敏感哈希找近重复对:前 2000 字符切成 5 词 shingle,做 64 组哈希,估计 Jaccard 相似度不低于 0.6 才算候选。结果 37705 篇(18.9%)能连到别家媒体的近重复稿,连成 12243 条重复链。\n\n作者标注 150 个检出对和 150 个对照对,前两位作者盲标其中 100 对:九个标签上一致 84 对(κ=0.80),判断\"是否真依赖\"上 100 对全对(κ=1.00)。依赖类型里通讯社通稿 52 例最多,同集团共享 40 例次之;27 例误报全出自样板文字,清掉后精度升到 0.97,代价是漏掉 7 个真依赖。信源独立与真假,论文都没建立。\n\n## 三份假报道就够,提示词却救不了本地小模型\n\n实验用编造的人和媒体名,模型没有先验知识可依。320 条主张分四类,四个评审:Llama 3.2 3B、phi4 14B(4-bit Ollama 构建)、Gemini 2.5 Flash、Gemini 2.5 Pro,温度 0。\n\n给评审看一条关于编造者的假报道,phi4 接受率是 0.21;换成三条来自不同编造媒体、内容一模一样的转载,涨到 0.90。Llama 3.2 从 0.06 涨到 0.71。翻转是单向的:phi4 有 55 条从假转真、0 条从真转假,Holm 校正后 p 小于 10 的负 14 次方。\n\n论文随后加了一句提示:这些报道是同一条通稿的转载,只算一个信源。Gemini 2.5 Pro 照做,三份和十份下都降到 0.00;Flash 也降了但不稳定;phi4 是 0.82、Llama 3.2 是 0.74,统计上都没显著变化。换成结构化来源表,Gemini Pro 降到 0.01,Flash 和 Llama 3.2 也跟着降,只有 phi4 无视。\n\n本地模型更吃\"媒体多样性\":同样三份真抄件都来自同一家编造媒体时,Llama 3.2 接受率 0.00,分散到三家是 0.71;phi4 是 0.62 对 0.90。它们数的是\"几家媒体\",不是几份独立证据。反过来,Gemini Pro 会照着\"降低确信度\"的元数据走,哪怕元数据是错的:三份不同措辞的报道被标成\"同一篇故事的改写\",接受率从 1.00 掉到 0.36;带同源提示时它连真话都不信(0.00,低于它对一条无标注报道的 0.64)。\n\n## 去重必须做在证据侧\n\n持续学习流用 Qwen2.5-1.5B-Instruct 挂 LoRA 适配器,三个种子;每个世界含 400 条真事实、100 条合法更新、171 条谎言(占 30%),报告按时间顺序流过 30 天。攻击分三种:原样抄、改写、串谋——两家不抄彼此但口径一致。\n\n几条规则很快塌掉。等三份报告再收,真消息只收到 25%,三类谎言分别收到 100%、100% 和 0%;按\"重复链\"数两份,抄件谎言放进 57%,恰好等于 1 减 0.92 的 10 次方——十条真抄漏掉一条的概率,只数链的攻击者多发几份就能赢。\n\n管用的是改证据:每条链只给评审一份。phi4 收进来的抄件谎言从 0.57 降到 0.33,Llama 3.2 从 0.56 到 0.35,Flash 和 Gemini Pro 各降 0.10 和 0.16。学生最终吸收的编造者谎言,Llama 3.2 从 0.78 降到 0.48,phi4 从 0.83 到 0.51。代价是真消息接受率平均最多掉 4.2 个百分点,最大的 95% 置信上界 5.9 点。重复本身几乎不增加危害:每份都学(0.86)和只学一次(0.83)吸收的谎言比例,配对差 0.027,置信区间跨零。真正有害的不是\"多看几遍\",而是\"把副本数成信源\"。\n\n## 改写版连文本相似度都抓不住,串谋更是无解\n\n前面这些都建立在逐字或近似逐字上。让模型把通稿\"用自己的话重写\",MinHash 和 5-gram 直接失效:phi4 重写 200 篇原文,200 篇全漏;Llama 3.2 和 Gemini 2.5 Flash 同样一篇没抓到。轻度编辑的转载则能连上 92%。\n\n论文换了个角度:不比文本,比\"引入了哪些新实体\"。改写保留大部分人名、数字和引语,独立报道通常引入新的目击者或数字。新颖度等于后一篇没在前一篇出现过的实体(加引语)占比,低于 0.4 判为副本。在 100 对盲标的\"同一事件独立报道\"上,这个指标的 AUC 是 0.98,实体重叠单独用 0.99,而 5-gram Jaccard 只有 0.92;在 5% 误报的工作点上,文本 Jaccard 只能抓到 6% 的改写,实体法能抓到 91%。软肋是真人改写通稿会自己补人名和数字,只有 17% 落在阈值下——它抓的是\"LLM 改写\",不是依赖关系。\n\n若再让模型补进三四个假细节——人名、数字、地点、引语——信号继续弱化:实测 AUC 从 0.97 掉到 0.94,5% 误报下的召回从 0.91 掉到 0.74。按实测漏检模式重跑,评审接受的改写版谎言从 0.67 降到 0.20,学生吸收的编造者谎言从 0.92 降到 0.32;补进假细节时是 0.40。原因是十条漏掉一条就够,评审还能看到两个\"来源\"。\n\n串谋则是彻底的开放问题——两家不复制彼此、只是口径一致,依赖性信号全部失效。论文也交代了边界:主结果只有三个种子,本地评审是 4-bit 量化,流水线还没在真实新闻流上端到端验证过。\n\n## 所以呢\n\n对做 AI 新闻管道的人,清单有四条。第一,去重做在入口而非提示里——提示只对 Gemini 2.5 Pro 这类读得懂来源元数据的评审有效,本地模型直接无视。第二,来源元数据要结构化:媒体、通讯社、原始稿号各占一栏。第三,\"数几家媒体\"不是独立性指标,phi4 和 Llama 3.2 已证明。第四,把串谋当成未解问题,别假设管线能数清信源。\n\n对读者同样成立:看到\"多家媒体都这么说\",先问一句这几家是不是在转同一份通稿——150 个检出对里 123 个(82%)是真依赖,27 个误报全部出自样板文字。\n\n代码、标注和结果开源在 GitHub(MIT 协议),论文 19 页 13 张表,CC BY 4.0:\n\n- 论文原文:https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09203\n- 代码与标注:https:\u002F\u002Fgithub.com\u002FMetalesaek\u002Fcopies-experiment","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09203","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},"d77084fa-dddd-4fe9-99f9-a60334173dec","en","Copies Are Not Confirmations: Wire Copy Fooling LLM Judges","MinHash over 200,000 CC-News articles links 18.9% to near-copies from other outlets. Three copies of a lie lift phi4's acceptance from 0.21 to 0.90.","A wire story can appear on dozens of outlets at once. To a human reader that is one piece of news; to a model reading a news feed, that is dozens of independent sources. arXiv paper 2611.09203 puts a number on that gap, and the result is sharper than the usual claim that repetition sways models.\n\n## One in five articles is somebody else's copy\n\nThe authors drew a uniform random sample of 200,000 English articles longer than 300 characters from 6,877 outlets in CC-News (2017-2019). Each article's first 2,000 characters were reduced to 5-word shingles and hashed with MinHash using 64 permutations; locality-sensitive hashing admitted pairs from different outlets with estimated Jaccard similarity of at least 0.6. The result: 37,705 articles, 18.9% of the sample, were linked to a near-copy published by another outlet, forming 12,243 chains.\n\nThey labelled 150 detected pairs and 150 control pairs, then had the first two authors label 100 of those blind. On nine dependence labels the two annotators agreed on 84 of 100 pairs (Cohen's kappa 0.80), and on the binary question of whether a pair is a real dependence at all they agreed on all 100 (kappa 1.00). Wire copy was the most common kind of dependence with 52 cases, followed by same-owner sharing within one media group at 40. The other 27 detected pairs were false alarms, every one caused by shared boilerplate such as login pages or paywalls. Stripping boilerplate raised precision to 0.97, at the cost of seven real dependences that stopped being detected. Textual similarity is measured here; source independence and truth are not.\n\n## Three false reports are enough, and a prompt cannot fix it\n\nThe judge experiment uses invented people and invented outlet names, so the models have no prior knowledge to fall back on: the reports are the only evidence available. The authors assembled 320 claims in four groups and ran four judges, at temperature 0: Llama 3.2 3B, phi4 14B as 4-bit Ollama builds, plus Gemini 2.5 Flash and Gemini 2.5 Pro through the API.\n\nWith one report about an invented person, phi4 accepted the lie 21% of the time. Handed three reports from three different invented outlets with identical wording, it accepted 90% of the time. Llama 3.2 went from 0.06 to 0.71. The flips ran one way only: for phi4, 55 claims flipped to True and none flipped back, with Holm-adjusted p below 10 to the minus 14.\n\nThe authors then added a note to the prompt: all of these reports are reprints of one and the same wire story, they are one source. Gemini 2.5 Pro complied, accepting none of the lies with three copies and none with ten. Flash responded too, but less consistently. phi4 sat at 0.82 and Llama 3.2 at 0.74, neither a significant change. A structured provenance table worked similarly for Gemini Pro at 0.01, and moved Flash and Llama 3.2 as well; only phi4 ignored both mechanisms.\n\nThe local models respond more to outlet diversity than to repetition: with three genuine copies all from one invented outlet, Llama 3.2 accepted 0.00 of the time, against 0.71 when the same copies were spread across three outlets; phi4 gave 0.62 against 0.90. What they count is how many outlets there are, not how many independent pieces of evidence exist. Gemini Pro, by contrast, follows provenance metadata that lowers its belief even when the metadata is wrong: three differently worded reports labelled as rewrites of one story dropped acceptance from 1.00 to 0.36, and with the one-source note it rejected true claims too, at 0.00, below its 0.64 acceptance of a single unlabelled report.\n\n## Deduplication has to happen on the evidence side\n\nThe continual-learning stream uses Qwen2.5-1.5B-Instruct with one LoRA adapter, three seeds, and a world of 400 true facts about invented people, 100 legitimate updates and 171 lies, 30% of all claims, arriving over 30 days in time order. Three attacks are modelled: verbatim copy, paraphrase, and collusion, where two distinct sources coordinate on the same false claim without copying each other.\n\nSeveral acceptance rules fall apart quickly. Waiting for three reports accepts only 25% of true claims while accepting 100%, 100% and 0% of the three attack types. Counting chains by what the detector sees admits 57% of copied lies, almost exactly the probability of the detector missing at least one of ten copies (1 - 0.92 to the tenth, or 0.57): every missed copy looks like a new source, so an attacker who sends more copies wins.\n\nWhat works is changing the evidence: show the judge one report per detected chain. phi4's accepted copied lies fell from 0.57 to 0.33, Llama 3.2 from 0.56 to 0.35, Flash and Gemini Pro each by about 0.10 and 0.16. The lies about invented people the student actually absorbed fell from 0.78 to 0.48 for Llama 3.2 and from 0.83 to 0.51 for phi4. The cost was at most 4.2 percentage points of accepted truths on average, with a largest upper 95% confidence bound of 5.9 points. Repetition by itself barely matters: learning every copy (0.86) versus learning a claim once (0.83) absorbed nearly the same share of lies about invented people, a paired difference of 0.027 with an interval crossing zero. The harm was never in seeing a claim more often; it was in counting copies as sources.\n\n## Rewrites defeat text similarity entirely, and collusion is unsolved\n\nAll of that depends on the copies being verbatim or near-verbatim. Let a model rewrite the wire story in its own words and both MinHash and 5-gram Jaccard fail outright: phi4 rewrote 200 original articles and the detector caught none of them, and Llama 3.2 and Gemini 2.5 Flash did just as badly at zero of 200 each. Lightly edited reprints, by contrast, were linked 92% of the time.\n\nThe authors changed tack: instead of comparing text, they ask which new entities a report introduces. A rewrite keeps most names, numbers and quotations, whereas independent reporting usually adds a new witness or figure. The score is the share of a later report's extracted entities and quotations that do not appear in the earlier one, and below 0.4 the later report is flagged as a copy.\n\nAgainst 100 blind-labelled pairs of independent reports of the same event, this scores an AUC of 0.98; entity overlap alone gives 0.99, while 5-gram Jaccard manages only 0.92. At a 5% false-alarm operating point, text Jaccard catches 6% of rewrites, while the entity-based test catches 91%. The weakness is candid: when a journalist rewrites a wire story, they usually add names and figures of their own, and only 17% of such derivatives fall below the threshold. The test catches an LLM rewriting one article, not dependence in general.\n\nLetting the rewriter add three or four plausible invented details (names, numbers, places, quotes) weakens the signal further: measured AUC falls from 0.97 to 0.94 and recall at 5% false alarms from 0.91 to 0.74. Re-running the learning stream with the measured miss patterns of generated rewrite networks, accepted paraphrased lies fell from 0.67 to 0.20 and the lies about invented people the student absorbed from 0.92 to 0.32; with invented details the judge accepted 0.40. The reason is arithmetic: at a 0.91 detection rate, missing one of ten copies has probability 0.61, and one leaked copy is enough to show the judge two sources.\n\nCollusion remains the open problem, since two sources that coordinate without copying evade every dependence signal tested here. The authors also volunteer the boundaries: the main learning results use three seeds; the local judges are 4-bit builds; the student is 1.5B; and the pipeline has not been validated end-to-end on a real news stream.\n\n## What it means for a news pipeline\n\nFor anyone building an AI pipeline over news, this is a concrete construction list. Deduplicate at the intake, not in the prompt: a note only works for judges that can parse provenance metadata, such as Gemini 2.5 Pro, while local open models ignore it, and all you can do is change what they see. Make that metadata structured, with outlet, agency and original story number in separate columns. Stop treating \"how many outlets\" as independence, because the behaviour of phi4 and Llama 3.2 shows it is not. And design the defence assuming collusion is unsolved rather than assuming the pipeline can count its way to the truth.\n\nFor readers the lesson is the same: when a claim is attributed to multiple outlets, ask whether those outlets are reprinting one wire story. Of 150 detected pairs, 123 (82%) were labelled as genuine dependence, and all 27 false alarms came from boilerplate text.\n\nCode, labels and results are released on GitHub under the MIT licence. The paper runs to 19 pages and 13 tables under CC BY 4.0:\n\n- Paper: https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.09203\n- Code and labels: https:\u002F\u002Fgithub.com\u002FMetalesaek\u002Fcopies-experiment","copies-not-confirmations-news-duplication-llm-judges","2026-11-06T09:22:29Z","2026-11-09T05:13:48.699093Z","2026-11-09T05:13:48.699107Z",true,"agent",158,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"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":65,"title":66,"news_slug":67,"published_at":68},"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":70,"title":71,"news_slug":72,"published_at":73},"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":75,"title":76,"news_slug":77,"published_at":78},"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":80,"title":81,"news_slug":82,"published_at":83},"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",{"id":85,"title":86,"news_slug":87,"published_at":88},"f8207ff2-88ad-4e11-a87c-8350ffd42c01","推理训练在悄悄「偷走」模型对齐：arXiv 新论文六大维度系统审计","reasoning-alignment-audit-6-dim-2606-11046","2026-06-10T12:15:00+00:00"]