Dongxi 东锡 NLP@dongxi_nlp
44AI 编辑部评分,满分 100
2026-08-15 03:45· 32分钟前
AI 导读

马东锡 NLP 解析 Claude 的 SynthID-Text 水印机制:模型前向传播后,对每个候选 token 进行一次带密钥的上下文相关“掷硬币”更新,再采样生成。持有密钥者可检测该统计模式。Anthropic 称此举为遵守欧盟 AI 法案,不影响输出质量、不增加成本,且水印无法追溯到具体个人或对话。

The SynthID-Text watermarking LLM generation is like:

model forward pass -> next-token logits ->temperature / top-k / top-p ---> keyed watermark probability update ->sample a token

For every plausible next token, Claude consults a secret, context-specific coin flip.

"Someone" holding the secret key can measure that pattern.

AnthropicWe've written an FAQ to answer some of the questions we've received about watermarking. In summary: • We're implementing watermarking to comply with the EU AI A...

来源:Dongxi 东锡 NLP · x.com

Dongxi 东锡 NLP · @dongxi_nlp · X·2026-08-15 03:45·32分钟前
AI 导读

马东锡 NLP 解析 Claude 的 SynthID-Text 水印机制:模型前向传播后,对每个候选 token 进行一次带密钥的上下文相关“掷硬币”更新,再采样生成。持有密钥者可检测该统计模式。Anthropic 称此举为遵守欧盟 AI 法案,不影响输出质量、不增加成本,且水印无法追溯到具体个人或对话。

The SynthID-Text watermarking LLM generation is like:

model forward pass -> next-token logits ->temperature / top-k / top-p ---> keyed watermark probability update ->sample a token

For every plausible next token, Claude consults a secret, context-specific coin flip.

"Someone" holding the secret key can measure that pattern.

AnthropicWe've written an FAQ to answer some of the questions we've received about watermarking. In summary: • We're implementing watermarking to comply with the EU AI A...

来源:Dongxi 东锡 NLP· x.com