Unverified50% confidenceFactExact time
统计学水印的核心原理是在大语言模型生成文本时，通过伪随机函数将候选token分为绿色列表和红色列表，略微提高绿色列表token的采样概率，通过z-test等统计检验方法检测
1
Sources
50%
Confidence
Long-term
Relevance
8/28/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMainstream LLM text watermarking schemes embed statistical signals by manipulating the token sampling probability distribution during model generation, rather than inserting special characters.76% similarUnverifiedKirchenbauer等人的水印算法在生成每个token时，根据前文的哈希值将词汇表划分为绿名单和红名单，并轻微提升绿名单token被选中的概率72% similarUnverified口语化采样通过在提示词中要求模型生成一组候选答案并显式说出每个答案对应的概率分布来提升多样性67% similarUnverifiedLarge language models use probability-based sampling strategies during text generation, where hyperparameters like temperature and top-p determine the degree of randomness in output.66% similarUnverifiedWhen generating each token, large language models perform probabilistic sampling over all possible next words in the vocabulary, rather than deterministically selecting a unique answer.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/810587API
curl https://kongchang.com/api/v1/knowledge/claims/810587MCP
get_claim(id=810587)