Unverified50% confidenceBenchmarkExact time
Self-Consistency通过多次采样推理路径后投票取多数答案,可进一步提升准确率约5-10个百分点
1
Sources
50%
Confidence
Long-term
Relevance
7/17/2026
First Seen
Sources
Related Claims
Unverified自我一致性(Self-Consistency)方法由Xuezhi Wang等人提出,在GSM8K等数学推理基准测试中将准确率提升了10%以上73% similarUnverified在歧义点的四类行为中,'先搜再问'成功率最高,平均通过率达93.4%,远高于'直接猜'的56.5%,'搜多次再猜'为51.5%70% similarUnverified测评题量与信度存在权衡:低于30题的『快测』因抽样不足信度偏低,但200题以上的专业量表易产生答题疲劳导致后半段随意作答,60-120题是兼顾信度与完成质量的区间69% similarUnverified在GSM8K数学推理基准上,Self-Consistency相比单次贪婪解码可将准确率提升约17个百分点68% similarUnverified研究数据表明,在数学竞赛级别问题上,启用深度推理的模型准确率可从不足30%提升至80%以上68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/546607API
curl https://kongchang.com/api/v1/knowledge/claims/546607MCP
get_claim(id=546607)