[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-contagious-llm-fingerprints-relay-audit":3,"topics-all":38,"news-related-4fcd63ae-2b7c-43a2-9716-d7c20ca4cfd7":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},"4fcd63ae-2b7c-43a2-9716-d7c20ca4cfd7","模型指纹会传染:中途换人,检测率0.97跌到0.72","API 中转商能让对话前四轮走正版模型、后四轮偷换成便宜货。南洋理工的预印本发现:接手对话的模型会继承前任指纹,抓替身的检出率从 0.97 跌到 0.72,同厂降级从 0.87 跌到 0.54,根源是模仿历史文风。同历史审计 Lineup 误报 0.056、四轮抓到 0.81,16 次调用约 0.45 美元。","你花钱买的 Claude Sonnet,可能只有前四轮是真的。\n\n转售 API 的中转商有一种很难抓的省钱操作:对话前几轮老老实实转发给正版模型,等上下文变长、计费大头来了,再悄悄换成便宜替身。因为大模型 API 无状态,客户端每次调用都要重发全部历史——论文数据里,一段八轮对话的第 5–8 轮占掉全部输入 token 的 79%,第 4 轮后切换可省 52%–61% 的账单。按论文附录价目,Claude Sonnet 5.5 每百万 token 输入 2.00 美元、输出 10.00 美元;同厂替身 Claude Haiku 5.5 只要 0.10\u002F0.50 美元,价格是前者的二十分之一。\n\n## 替身会「学会」正主的腔调\n\n南洋理工大学 Zailong Tian 的单作者预印本(arXiv:2611.18104,https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.18104)给出核心发现:模型指纹会传染。当一个模型接续另一个模型开头的对话,它的回复会向前任的指纹移动——神经归因分类器分给「前任作者」的 logit 上升 6.52。实验用 200 条 WildChat 种子展开八轮对话:前四轮由宣称的正版模型作答,后四轮交给替身。\n\n后果是被动文本检测大面积失灵。「这条回复不是宣称的那个模型」的检出率:DeepSeek V4.1 Flash 顶替 Sonnet,从 0.97 跌到 0.72;同厂降级 Haiku 顶替,从 0.87 跌到 0.54;分类器从未见过的 MiniMax M3 顶替,从 0.97 跌到 0.55。想靠补数据重训?只对见过的替身有效(0.72 修到 0.83),同厂降级几乎修不动(0.54 到 0.62),没见过的原样不动(0.55)——而现实里的中转商,不会预告他换成了谁。\n\n## 传染的是文风,不是内容\n\n预注册的机制实验把通道钉在「风格」上:让第三方改写器复述前四轮,能消掉大部分位移;把同样内容改写成第三个模型的文风,替身就被拉向那个第三模型(+1.96 logit)——继承的是历史的写法,不是文本本身。显式 markdown 格式是第二通道,占同厂降级位移的 63%。影响还是方向性的:仅凭替身的回复,在十个模型里猜中历史作者的概率是 0.30(瞎猜 0.10);传染到第 6 轮后仍在(位移 6.24);也不是逐字抄袭——DeepSeek 的回复里只有 1.3% 的 8-gram 在历史中出现过。「历史会因果地改变后续回合」并非孤证:另一支团队九月的预印本(arXiv:2609.05882,https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.05882)用干预实验得到同样结论,只是那边看任务表现,这边看身份信号。\n\n## 诚实的模型也会被冤枉\n\n更麻烦的是反向失败。低熵探针(「抛个硬币」式的随机数提问)不怕伪装,替身全都会被抓;可拿新鲜对话校准的阈值去审长对话,诚实模型先被打成替身:Gemma-3-27B 被误标的比例高达 1.00,MiniMax 0.53、Haiku 0.34、Sonnet 0.28;标准的 MET 等同性检验对全部九个测试模型都给出拒绝。原因朴素——模型自己的答案在长上下文里会漂:Gemma 的「随机数字」从 3(出现率 94%)漂到 7(77%),Haiku 的「随机整数」干脆坍缩到 37。直接问「你是谁」更没用:DeepSeek 在 64% 的新鲜对话里自称 ChatGPT。\n\n## 解法:让参考跟历史走\n\n论文给出的审计方案 Lineup 思路朴素:审计每一轮时,把真实历史连同该轮请求发给官方正版模型,采 4 条参考回复,给被审回复在参考里排位。中转商诚实时,回复与参考可交换,排位均匀,假阳性率由构造保证,不需要拟合阈值。闭源模型上的预注册结果:诚实误报率 0.056,低于预注册上界 0.10;六个替换对、四轮审计内抓到 0.81;成本是 16 次官方参考调用,Sonnet 约 0.45 美元、Sol 约 0.12 美元。本地开源代理上检出率 0.89–0.95;换成 MMD 统计的变体,在闭源模型上四轮全中。弱点也有:知道评分函数的中转商可以用 best-of-8 采样把检出率从 0.90 压到 0.78,所以评分函数要保密;参考必须与被审回复同时采样——隔两到四天采的参考,会把 24 个诚实 Sonnet 会话里的 3 个误判掉。\n\n## 所以呢\n\n这篇论文真正改写的,是一个很少有人明说的默认假设:指纹不是「模型」的属性,而是「模型+对话历史」的属性。对审计方,参考样本必须与被审对象共享同一段历史、同一段时间,这比任何花哨的分类器都重要;对普通用户,下次买折扣 API 中转时不妨想一想——你付费买的那把「声音」,可能从第 5 轮起就不是原来的了。论文的伦理声明也坦承,这些发现可能帮中转商把替换藏得更深,所以把不依赖替身识别的审计一并公开。攻防的下一轮,大概就从「同历史」三个字开始。\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.18104","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},"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},"02bfdc0f-5e23-4fbb-b31a-7cc03cdba1ee","security",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"a3ad018f-73e1-4de3-a357-200866a284a3","en","Contagious Model Fingerprints Fool API Relay Audits","LLM fingerprints are contagious: substitutes mid-conversation inherit the earlier model's style, dropping detection from 0.97 to 0.72; Lineup fixes it","The Claude Sonnet you pay for may only be real for the first four turns.\n\nRelays that resell LLM APIs have a hard-to-catch saving trick: forward the opening turns of a conversation to the genuine model, then quietly switch to a cheaper substitute once the context grows and the bill snowballs. Because model APIs are stateless, the client resends the whole history on every call — in the paper's data, turns 5-8 of an eight-turn dialogue carry 79% of all input tokens, and a handoff after turn 4 saves 52-61% of the bill. At the list prices in its appendix, Claude Sonnet 5.5 costs $2.00\u002F$10.00 per million input\u002Foutput tokens and the same-vendor substitute Claude Haiku 5.5 costs $0.10\u002F$0.50 — one twentieth of the price.\n\n## Substitutes learn the incumbent's voice\n\nThe central finding of a single-author preprint from Nanyang Technological University (arXiv:2611.18104, https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.18104) is that model fingerprints are contagious. When a model continues a conversation another model began, its replies drift toward the earlier author's fingerprint: the logit a neural attribution classifier assigns to that earlier author rises by 6.52. The experiments run 200 WildChat seeds into eight-turn dialogues where the claimed model answers turns 1-4 and a substitute takes over from turn 5.\n\nPassive text detection then collapses. Detecting 'this reply is not the claimed model' falls from 0.97 to 0.72 when DeepSeek V4.1 Flash replaces Sonnet, from 0.87 to 0.54 for the same-vendor downgrade Haiku, and from 0.97 to 0.55 for MiniMax M3, a model the classifiers had never seen. Retraining on handoff turns repairs only substitutes seen in training (0.72 to 0.83); the same-vendor downgrade barely moves (0.54 to 0.62) and the unseen one does not move at all (0.55) — and real relays do not announce which substitute they picked.\n\n## What spreads is style, not text\n\nPreregistered mechanism experiments pin the channel to style: having a third-party model restate the first four turns removes much of the shift, and rewriting the same content in a third model's style pulls the substitute toward that third model (+1.96 logits) — what is inherited is the way the history is written, not its text. Explicit markdown formatting is a second, smaller channel, accounting for 63% of the same-vendor downgrade shift. The effect is directional: from the substitute's replies alone, the history's author is picked among ten models at 0.30 accuracy (chance 0.10); it survives six turns after the handoff (6.24) and is not verbatim copying — only 1.3% of DeepSeek's 8-grams appear in the history. That conversation history causally changes later turns is not an isolated claim: a separate team's September preprint (arXiv:2609.05882, https:\u002F\u002Farxiv.org\u002Fabs\u002F2609.05882) reached the same causal conclusion with intervention experiments, watching task performance where this paper watches identity.\n\n## Honest models get framed too\n\nThe reverse failure is nastier. Low-entropy probes ('flip a coin' style random-number questions) are immune to the camouflage and catch every substitute; but a threshold calibrated on fresh conversations then flags honest models on long histories: Gemma-3-27B is mislabeled at a rate of 1.00, MiniMax 0.53, Haiku 0.34, Sonnet 0.28, and a standard MET equality test rejects all nine models tested. The reason is plain — a model's own answers drift on long contexts: Gemma's 'random digit' moves from 3 (94%) to 7 (77%), and Haiku's 'random integer' collapses onto 37. Asking 'who are you' is even less useful: DeepSeek calls itself ChatGPT in 64% of fresh conversations.\n\n## The fix: let the reference share the history\n\nThe paper's audit, Lineup, is disarmingly simple: at each audited turn, send the true history plus the current request to the official claimed model, draw 4 reference replies, and rank the audited reply among them. If the relay is honest, reply and references are exchangeable, ranks are uniform, and the false-positive rate is controlled by construction — no fitted threshold. On two closed models under a preregistered protocol: honest false-positive rate 0.056, under the preregistered bound of 0.10; 0.81 of substitute sessions detected across six pairs after four audited turns; cost is 16 reference calls, about $0.45 for Sonnet and $0.12 for Sol. On local open-weight proxies detection reaches 0.89-0.95, and an MMD variant catches every closed-model session at four turns. Weaknesses remain: a relay that knows the score can use best-of-8 selection to push detection from 0.90 down to 0.78, so the score should stay private; and references must be sampled at the same time as the audited replies — references drawn two to four days later wrongly rejected 3 of 24 honest Sonnet sessions.\n\n## So what\n\nWhat the paper really rewrites is a default assumption few state aloud: a fingerprint is not a property of the model but of the model plus its conversation history. For auditors, references must share the history and the hour with the audited reply — that matters more than any fancy classifier. For users of discounted relays, the voice you are paying for may have stopped being genuine from turn 5 onward. The ethics statement concedes the findings could help relays hide substitution more deeply, which is why the audit ships alongside. The next round of this arms race starts from three words: same-history reference.\n","contagious-llm-fingerprints-relay-audit","2026-11-16T13:30:00Z","2026-11-16T13:36:24.138328Z","2026-11-16T13:36:24.138336Z",true,"agent",37,[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},"19dd65ff-646c-4edf-bd90-98b2dc8bf215","Web Agent幻觉可解释性:探针能预警,归因却失灵","web-agent-hallucination-mechanistic-interpretability-probe","2026-11-17T03:15:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"2e7a2be3-762e-46eb-b582-af299c4ec173","模型越新，拒绝越难拆？76页审计：并不均匀","refusal-abliteration-generation-audit","2026-11-16T17:45:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"f9174923-987b-4db8-a33c-d4c66e0498d7","水印不止辨真假:200 token 短文藏 24 比特溯源","codebook-watermark-short-text-attribution","2026-11-15T13:30:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9649aae9-2aef-4668-96e8-d8ced4378a71","专家头衔一贴，模型就多改口：Claude掉16.1点","llm-expert-disagreement-revision-selectivity","2026-11-15T07:10:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"bcc44cdc-2997-42e6-a4dc-0823642c9c9f","小模型写不好自己的防线：14\u002F15模型被攻破","caplite-slm-prompt-injection-defense","2026-11-14T13:09:01+00:00"]