Unverified50% confidenceFactExact time
The research team analyzed item-level response data from 47 publicly released models and 74 models fine-tuned with known contamination doses.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRecent research indicates that some models perform 10–15 percentage points higher on known contaminated benchmark subsets than on uncontaminated subsets.64% similarUnverifiedAccording to Cohen's d effect size criteria, detecting a medium-sized effect typically requires at least 30–50 samples per group.63% similarUnverifiedMeituan's technical team has released the General 365 inference evaluation benchmark, testing 26 mainstream large models, with the majority failing to reach the passing score of 60.63% similarUnverifiedChinchilla论文通过对超过400个不同规模模型的系统性实验得出结论61% similarUnverifiedScale AI的研究团队发现，部分模型在被污染的基准上的表现比在新编写的同等难度题目上高出10-15个百分点61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/900507API
curl https://kongchang.com/api/v1/knowledge/claims/900507MCP
get_claim(id=900507)