Unverified50% confidenceBenchmarkExact time
在留出测试集上，FakeSpotter对短文本取得0.788的宏观F1分数，对长文本取得0.793的宏观F1分数
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
Unverified早期在Opus模型上测试精确字符串识别率为15分之0，而Fable 5发布后同样测试变成15分之1366% similarUnverified在512-4096上下文长度测试中，将草稿长度从4字符拉满至16字符，整体推理延迟仅增加0.2%-1.3%64% similarUnverified仅停用词移除配置可实现token削减29.6%，BERTScore-F1保持在0.91364% similarUnverifiedQwen 3.6 Max中性模式下歧义检测F1仅16%，平均每任务只问0.07个问题，但一旦提问澄清准确率高达94.7%63% similarUnverified研究团队在15个长文本任务上对现成模型进行了零样本评估，测试对象包括Gemma-4-31B和Qwen-3.6-27B60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/931147API
curl https://kongchang.com/api/v1/knowledge/claims/931147MCP
get_claim(id=931147)