Unverified50% confidenceFactExact time
ContextualJudgeBench tests whether judgments are affected by position bias by presenting items in both orders, counting a case as passed only when both orders are judged correctly, which reduces the random baseline from 50% to 25%.
1
Sources
50%
Confidence
Long-term
Relevance
10/12/2026
First Seen
Sources
Related Entities
Related Claims
Unverified多次采样取多数票或自洽性方法可在推理层降低单次随机性对最终结果的影响67% similarUnverifiedThe pairwise verifier requires preference for the negative twin under both presentation orders, A vs B and B vs A, in order to suppress the positional bias of the LLM verifier.65% similarUnverified浮点运算不满足结合律，不同线程束的执行顺序差异会导致数值结果出现细微偏差64% similarUnverified决策树的分裂准则（信息增益、基尼不纯度、均方误差）本质上是排序统计量的函数，只关心阈值两侧的类别比例而非样本间距离63% similarUnverifiedWriting only one position or conditionally independent positions per step can theoretically match the distribution exactly, while writing more dependent positions causes errors to accumulate.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1001725API
curl https://kongchang.com/api/v1/knowledge/claims/1001725MCP
get_claim(id=1001725)