“On the technical front there's a good paper here: A Watermark for Large Language Models, and a nice very approachable video explaining it here: [LINK].”
A Watermark for Large Language Models
用绿色词表软性偏置采样,给 LLM 文本加水印
提出面向专有语言模型的水印框架:在生成每个词之前随机选出一组绿色 token,采样时软性地提高使用绿色 token 的概率,从而在文本中嵌入人眼不可见、但算法可检测的信号。水印对文本质量影响可忽略,检测无需访问模型 API 或参数,并配有带可解释 p 值的统计检验和信息论分析框架。作者在 OPT 系列的数十亿参数模型上做了测试,并讨论了鲁棒性与安全性。
评论者常把它称为红绿词表方案,是讨论如何识别 AI 生成文本时的标准出处,同时也有人指出这种水印很容易被绕过。
HN 书评
推荐在前,同类按时间 · 每篇最多保留 20 条,统计用全部提及“some interesting work using steganography to mark generated text”
“Here's a decent paper on it.”
“A really good paper about this idea was just awarded outstanding paper at ICML - A Watermark for Large Language Models.”
“is not an example of this because of how trivial it is to defeat.”
“I think watermarking is a better solution. There is a paper on it:”
看全部 13 条 →
“I wonder if they have a watermark built in that allows them to detect text generated by Llama easily”
“text watermarking works pretty well! A Watermark for Large Language Models”
“closer to the original "red token, green token" paper: A Watermark for Large Language Models”
“explaining a paper by John Kirchenbauer, Jonas Geiping et al. that described a method for watermarking LLM output like this.”
“A token is hashed and used to seed a random number generator, which produces the red list for the token after it.”
“the easiest for people to understand is the red/green scheme by Kirchenbauer et al., which does technically distort the logits (but I'd argue only in cases where you wouldn't notice it anyway).”
“A reasonable guess about the algorithm is 'A Watermark for Large Language Models'.”
- A Watermark for Large Language Models ↗lnyan · 2023-01-25 · 1 分
- A Watermark for Large Language Models ↗tosh · 2023-01-25 · 2 分
- A Watermark for Large Language Models ↗headalgorithm · 2023-01-26 · 3 分
- A Watermark for Large Language Models ↗alexmolas · 2023-02-01 · 1 分
- A Watermark for Large Language Models [pdf] ↗qingdao99 · 2023-02-17 · 2 分
- A Watermark for Large Language Models ↗EndXA · 2023-11-15 · 1 分