Unverified50% confidenceBenchmarkExact time
在Qwen-3.5-4B-Instruct模型上，前两条主轴与人类自发请求的风格维度匹配精确率达72.8%，宏平均召回率43.6%
1
Sources
50%
Confidence
Long-term
Relevance
9/18/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在Qwen-3.5-4B-Instruct上，75.6%的有效性评分认为主轴两极生成结果准确对应其标签，标注者相邻一致率高达90.9%73% similarUnverifiedQwen3.7 Max的输出质量高度依赖提示词质量，同一模型在不同提示词下的性能差异可高达30%-50%73% similarUnverified在六个Qwen系列模型上进行匹配帧数（matched-frame）的对照实验，字幕方案在帧数对齐条件下平均准确率提升3.22分和2.55分72% similarUnverified微调后的Qwen3.5-4B模型在BU Bench V1基准上准确率从15%提升至53%69% similarUnverified根据Artificial Analysis数据，Qwen3.8B在600多个参与排名的模型中位列第26名68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/931130API
curl https://kongchang.com/api/v1/knowledge/claims/931130MCP
get_claim(id=931130)