Unverified50% confidenceOpinionExact time
人类主观投票能捕捉难以量化的质量差异，成为FID、CLIP Score等自动化指标的重要补充
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
Verified自我一致性（self-consistency）采样通过并行生成多个候选答案并多数投票来提升可靠性，无需外部验证器67% similarUnverified能清晰阐述为什么选A而不选B的候选人才能通过Staff级别的评估67% similarUnverified锚定样本技术在评估提示词中加入已由人类专家标注明确分数的示例，覆盖评分量表各个档位，可帮助裁判模型准确理解评分标准65% similarUnverifiedCoT-SC利用大模型输出的随机性，对同一问题多次调用后对结果投票取多数，能显著提升推理结果的可靠性65% similarUnverified标注歧义性对应机器学习中的贝叶斯最优误差率概念，可通过让多名标注者独立标注并计算标注者间Dice系数来估算性能天花板64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/913654API
curl https://kongchang.com/api/v1/knowledge/claims/913654MCP
get_claim(id=913654)