Unverified50% confidenceBenchmarkExact time
MMLU评分从GPT-3.5的约70%跃升至GPT-4的86%，再到后续模型逼近90%以上
1
Sources
50%
Confidence
Long-term
Relevance
8/19/2026
First Seen
Sources
Related Claims
UnverifiedGPT-4在2023年初于MMLU上取得86.4%的成绩80% similarUnverified在 MMLU、HumanEval、GSM8K 等主流基准上，开源模型已达到 GPT-4 早期版本 85-95% 的水平77% similarUnverifiedGPT-5.5在GPQA博士级科学推理基准上从78.5%提升到85.6%74% similarUnverifiedGPT-4在LeetCode竞赛题上的通过率超过80%73% similarUnverified顶级自回归模型如GPT-4o、Claude 3.5在GSM8K上准确率已超过95%，导致基准饱和72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/774422API
curl https://kongchang.com/api/v1/knowledge/claims/774422MCP
get_claim(id=774422)