[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-codebook-watermark-short-text-attribution":3,"topics-all":38,"news-related-f9174923-987b-4db8-a33c-d4c66e0498d7":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},"f9174923-987b-4db8-a33c-d4c66e0498d7","水印不止辨真假:200 token 短文藏 24 比特溯源","检测水印只答是不是 AI 写的,溯源答哪个账号流出。波兰科学院与 CSIRO 团队提出码本式多比特水印:24 比特会话标识编进 200 token,提取端只要密钥和分词器查码本。论文实测 Llama 2 7B 恢复率 99.8-99.9%,指令模型 127 token 掉到 61.6%,词插入剩 57.3%。","AI 文本水印在公众语境里主要是个「检测」工具:给一段话,判断是不是模型写的。但平台真正难办的是另一件事——一段被泄出去的 AI 短文本,到底是从哪个账号、哪次会话流出去的。检测回答「是不是」,溯源回答「是谁」。11 月 12 日挂上 arXiv 的论文《Look It Up: Multi-bit LLM Watermarking via Codebook Search》,来自波兰科学院计算机科学研究所与澳大利亚 CSIRO 的三位作者,把短文本溯源往工程上推了一步:200 个 token 的生成文本里,稳定藏进 24 比特的身份信息。\n\n## 思路:身份不逐位解码,改成查表\n\n多比特水印在短文本上一直难做:载荷位数一多,要么靠长文本慢慢积累信号,要么解码端代价高。这篇论文把编码问题变成查表问题。嵌入端先把 n 比特身份映射成更长的 u 比特码字(u 大于 n,冗余就是纠错空间),生成时按码字逐位施加偏置:词表按前一个 token 和密钥的哈希分成绿红两半(沿袭 Kirchenbauer 2023 的绿名单思路),当前位是 1 就给绿半边加 logit 偏置(默认强度 6),是 0 就偏置红半边,默认贪心解码。提取端只要密钥和分词器:重新分词、逐 token 读回一个带噪比特串,再去预生成的秘密码本里做近邻匹配,汉明距离或编辑距离都可以。\n\n码本是 GF(2) 上的随机线性码:矩阵和偏移由密钥经 SHA-256 派生,没有密钥就重建不了。论文算了几笔账:n 取 24、u 取 127 时,码字之间的汉明距离下限是 35,汉明检索可纠 17 位以内的翻转;任意两个身份撞码字的概率上界低到 2 的负 80 次方量级。\n\n## 数字:基座模型近乎全中,指令模型露短板\n\n按论文实测,Llama 2 7B 在 200 token 文本上,16-24 比特载荷的恢复率是 99.8-99.9%;同一张对比表里,五个先前方法在同档成绩从 30.4% 排到 99.6%。换到 127 token 短文本、24 比特载荷:四个基座模型(Llama 2 7B、Llama 3.1 8B、Llama 3.2 3B、Mistral 7B)落在 97.5-100%,两个指令模型明显掉队,码本变体在 127 token 掉到 61.6%,200 token 才回到 90.6% 以上。论文的解释很直白:指令模型的输出分布低熵而尖峰,偏置信号压不过分布。\n\n速度是这套方案敢谈落地的原因:n 取 24 时码本约有 1680 万个码字,精确暴力检索单条文本 4.1-75.5 毫秒(汉明 CPU 到 Levenshtein GPU,AMD EPYC 7302 加一块 RTX 5090 实测)。鲁棒性方面,Levenshtein 变体在删掉 10% 的词后保住 95.2% 恢复率,同量级段落改写后保住 98.9%;词插入是硬伤,只剩 57.3%,同义替换也掉到 76.4%。区分水印与非水印文本,三个变体的 AUROC 都在 0.998 以上。\n\n## 代价:困惑度翻倍,码本吃内存\n\n没有免费的水印。论文自己的质量评估里,200 token 文本的中位困惑度从 4.22 涨到 9.58(评测模型 Gemma 4 31B),翻了一倍多——人未必读得出来,统计瞒不住。存储上,n 取 24、u 取 127 的码本约 256MB,索引还要再各来一份;32 比特载荷在这个框架下要 64GB 量级,更长的身份得拆块处理。还有一个结构性限制:嵌入需要访问 logits,这套方案只能由跑模型的一方来做——它是平台自证工具,不是第三方取证工具;密钥一旦泄露,整套溯源体系作废。\n\n## 所以呢\n\n这篇论文的价值不在又一个高恢复率,而在把「短文本能不能带身份」算成一笔清楚的工程账:毫秒级检索、百兆级码本、可量化的质量损耗,以及明码标价的短板(指令模型、词插入、强改写)。当监管和平台开始认真讨论 AI 内容的可追溯,这类账会被反复翻出来对。反过来说,服务方若没有动力嵌入,再漂亮的码本也只是论文里的数字——溯源水印的根本问题是激励,不是编码。\n\n平台会愿意给自己的输出「埋身份证」吗?评论区聊聊。\n\n论文:[arxiv.org\u002Fabs\u002F2611.16890](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16890) 代码:[github.com\u002Fjaroslawjanas\u002Fllmw-codebook](https:\u002F\u002Fgithub.com\u002Fjaroslawjanas\u002Fllmw-codebook)\n","https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16890","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},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"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},"e5248d93-0942-49ef-bddc-bdf3d862060c","en","Codebook watermarks hide 24-bit IDs in 200-token texts","A 24-bit ID hides in 200 tokens; codebook search recovers it on 99.8-99.9% of Llama 2 7B texts. Instruct models drop to 61.6% at 127 tokens.","Watermarking for LLM text is usually framed as detection: given a passage, decide whether a model wrote it. The harder production question is attribution. When a short AI-generated text leaks, which account or session did it come from? Detection answers whether; attribution answers who. On 12 November 2026 a paper posted to arXiv, Look It Up: Multi-bit LLM Watermarking via Codebook Search, by three authors from the Institute of Computer Science of the Polish Academy of Sciences and Australia's CSIRO, pushes short-form attribution a step closer to engineering practice: 24 bits of identity, stably embedded across 200 generated tokens.\n\n## The scheme: stop decoding bit by bit and look it up\n\nMulti-bit watermarking has struggled on short texts: with more payload bits you either wait for a long text to accumulate signal or pay heavily at the decoder. This paper turns coding into a table lookup. At embed time an n-bit identity is mapped to a longer u-bit codeword (u greater than n; the redundancy is the error-correction budget). During generation the vocabulary is split into green and red halves by a hash of the previous token and the secret key, following the green-list idea of Kirchenbauer et al., 2023; a logit bias (default strength 6) is added to the green half when the current bit is 1 and to the red half when it is 0, with greedy decoding by default. Extraction needs only the secret key and the model's tokenizer: re-tokenize the text, read back one noisy bit per token, then find the nearest codeword in a precomputed secret codebook, by Hamming or Levenshtein distance.\n\nThe codebook is a random linear code over GF(2): the matrix and offset derive from the key via SHA-256, so the codebook cannot be rebuilt without it. The paper does the engineering arithmetic: at n = 24 and u = 127 the minimum Hamming distance between codewords is 35 — Hamming search corrects up to 17 flipped bits — and the bound on two identities colliding onto the same codeword is on the order of 2 to the minus 80.\n\n## The numbers: base models near-perfect, instruct models expose the weak spot\n\nPer the paper's own runs, on 200-token texts from Llama 2 7B, 16-24 bit payloads are recovered at 99.8-99.9%; in the same comparison table the five prior methods span 30.4% to 99.6% over that payload range. Move to 127-token short texts with a 24-bit payload and the four base models (Llama 2 7B, Llama 3.1 8B, Llama 3.2 3B, Mistral 7B) land at 97.5-100%, while the two instruction-tuned models fall behind: the codebook variants bottom out at 61.6% at 127 tokens and climb back above 90.6% at 200. The explanation is blunt: low-entropy, peaked output distributions overpower the bias signal.\n\nSpeed is why this scheme can talk about deployment: at n = 24 the codebook holds roughly 16.8 million codewords, and exact brute-force search takes 4.1-75.5 ms per text (Hamming on CPU to Levenshtein on GPU, measured on an AMD EPYC 7302 with one RTX 5090). On robustness, the Levenshtein variant keeps a 95.2% match rate after 10% of words are deleted and 98.9% after paraphrasing of a span of the same size; word insertion is the hard case at 57.3%, and synonym replacement drops to 76.4%. Separating watermarked from non-watermarked text, all three variants reach AUROC above 0.998.\n\n## The costs: perplexity more than doubles, and the codebook eats RAM\n\nThere is no free watermark. In the paper's own quality evaluation, median perplexity on 200-token Llama 2 7B texts rises from 4.22 to 9.58 (evaluated with Gemma 4 31B) — more than doubling; human readers may not notice, statistics will. On storage, the n = 24, u = 127 codebook takes about 256 MB with indexes on top; a 32-bit payload would need about 64 GB in this framework, so longer identities must be split into blocks. And a structural limit: embedding requires access to logits, so only whoever runs the model can do it — a platform self-certification tool, not third-party forensics. Leak the key and the whole attribution scheme is dead.\n\n## So what\n\nThe value here is not one more high recovery rate; it is a legible engineering ledger for short-text attribution: millisecond lookup, hundred-megabyte codebooks, quantified quality loss, and clearly priced weak spots (instruct models, insertions, heavy rewrites). As regulators and platforms get serious about traceable AI output, these are the accounts that will get audited. The flip side: if the party running the model has no incentive to embed, the prettiest codebook stays a number in a paper — the fundamental problem in attribution watermarking is incentives, not coding.\n\nWould a platform volunteer to plant ID cards in its own outputs? That question is worth more than any recovery rate.\n\nPaper: [arxiv.org\u002Fabs\u002F2611.16890](https:\u002F\u002Farxiv.org\u002Fabs\u002F2611.16890) Code: [github.com\u002Fjaroslawjanas\u002Fllmw-codebook](https:\u002F\u002Fgithub.com\u002Fjaroslawjanas\u002Fllmw-codebook)\n","codebook-watermark-short-text-attribution","2026-11-15T13:30:00Z","2026-11-15T13:22:49.675163Z","2026-11-15T13:22:49.675171Z",true,"agent",61,[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},"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":65,"title":66,"news_slug":67,"published_at":68},"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","contagious-llm-fingerprints-relay-audit","2026-11-16T13:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"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",{"id":75,"title":76,"news_slug":77,"published_at":78},"af484c32-d6f7-406e-86a0-09dc2cdabbc0","零宽字符藏在回复和邮件里:模型偷偷记下你的性别年龄","qwen-llama-synthesis-smuggling-privacy-leak","2026-11-14T03:05:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"ef87ba13-75d5-41ff-a6d0-0a6e8b2cf30d","偏好标签三成翻错:PORT照拿67.6%胜率","port-token-reweighting-noisy-preference","2026-11-10T17:30:00+00:00",{"id":85,"title":86,"news_slug":87,"published_at":88},"92267597-eb39-4323-a5db-f7db3821f392","安全对齐被反用:一条假差评就能改掉智能体的选择","voluntary-servitude-agent-safety-attack","2026-11-05T01:05:00+00:00"]