Unverified50% confidenceFactExact time
2024年研究显示，经过专门训练的小型改写模型可在保持95%以上语义相似度的同时将水印检测率降至接近随机水平
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn 2023, researchers demonstrated that contaminating less than 1% of training data is sufficient to alter model behavior under specific trigger conditions72% similarUnverifiedResearch indicates that through one round of high-quality rewriting, the watermark signal can be weakened by 60%–90%.68% similarUnverified实验数据显示，经过5-10轮迭代训练后，模型在低频词汇和长尾知识上的覆盖率下降超过40%，而输出文本在表面流畅性上几乎没有可察觉的退化67% similarUnverifiedA 2024 study indicates that even when AI-generated content accounts for only 10–20% of training data, measurable quality degradation can occur after multiple generations of training.66% similarUnverified参数量在1亿以下的轻量模型经领域数据微调后可达95%以上意图分类准确率，单次推理延迟控制在5-20ms以内66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/793821API
curl https://kongchang.com/api/v1/knowledge/claims/793821MCP
get_claim(id=793821)