Verified65% confidenceFactExact time
Multiple academic studies have shown that a considerable proportion of the high scores achieved by mainstream models such as GPT-4 and LLaMA on leaderboards including MMLU, HumanEval, and GSM8K stem from overlap between training set and test set content.
3
Sources
65%
Confidence
Long-term
Relevance
7/9/2026
First Seen
Sources
The Truth Behind AI Benchmark Leaderboards: Real-World Performance Is the True Moat
bilibili卢菁博士_北大AI博士后7/7/2026
Related Claims
UnverifiedGPT 系列、Llama 等模型的核心能力绝大部分来自预训练而非后续微调阶段77% similarUnverifiedThe GPT-4o, o1, and o3 series have continuously set records on specific benchmarks (such as MATH, HumanEval, and MMLU), yet many developers find it difficult to perceive improvements of the same magnitude in everyday use.74% similarUnverified训练阶段化方法包括GPT-4等模型采用的渐进式sequence length增长、Chinchilla论文启发的动态batch size调度、μP参数化的分层学习率缩放73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/355048API
curl https://kongchang.com/api/v1/knowledge/claims/355048MCP
get_claim(id=355048)