Unverified50% confidenceBenchmarkExact time
早期的Codex模型在HumanEval测试上的通过率约为28.8%
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
Unverified2023 年一项学术研究发现在特定提示条件下 Codex 模型约 1% 的输出与训练数据存在逐字匹配66% similarUnverifiedOn professional code benchmarks such as HumanEval, the performance of top models has approached or even surpassed average human-level performance.59% similarUnverifiedIthaca模型缺失文字预测准确率约62%，人类专家独立完成时约为25%，地理来源推断约71%，年代估算误差中位数约30年58% similarUnverified修复前模型每约29次下一token预测中约有1次的选择与数学上应有结果不同56% similarUnverified审计数据显示 Codex 有 40% 的时候会提交 draft（草稿）PR，其他模型则很少这样55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/793867API
curl https://kongchang.com/api/v1/knowledge/claims/793867MCP
get_claim(id=793867)