Unverified50% confidenceBenchmarkExact time
使用思维链（CoT）提示要求裁判先分析再打分可将裁判与人工标注的一致性提升约7-12%
1
Sources
50%
Confidence
Long-term
Relevance
7/16/2026
First Seen
Sources
Related Claims
Unverified在采信LLM裁判评分前必须用人工标注对其进行校准，缓解偏差手段包括交换候选顺序取平均、使用CoT提示、多裁判集成78% similarUnverifiedCoT prompting techniques have improved accuracy by tens of percentage points on tasks such as mathematical reasoning and logical analysis.71% similarUnverifiedIn the hat pickup system test, Claude correctly implemented overlap detection, input enable/disable logic, and attaching the hat to the character's head socket, earning a score of 8–9.67% 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67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/530941API
curl https://kongchang.com/api/v1/knowledge/claims/530941MCP
get_claim(id=530941)