Unverified50% confidenceFactExact time
Scale AI在2024年的分析显示，部分模型在MMLU上的分数膨胀可能高达5-10个百分点，因测试集已泄露到训练数据中
1
Sources
50%
Confidence
Long-term
Relevance
8/5/2026
First Seen
Sources
Related Claims
UnverifiedA 2024 study indicates that even when AI-generated content accounts for only 10–20% of training data, measurable quality degradation can occur after multiple generations of training.78% similarUnverifiedSince the second half of 2024, the score improvements of mainstream large models on standard benchmarks such as MMLU and HumanEval have narrowed noticeably.76% similarUnverifiedBy mid-2024, top models had generally exceeded scores of 85–90% on MMLU, causing the benchmark's discriminative power to drop sharply.73% similarUnverifiedAI benchmark suites like MMLU, HumanEval, and MATH may produce inflated scores due to benchmark contamination, where models have indirectly encountered test data during training.72% similarUnverified2022年Hoffman等人的Chinchilla论文重新校准了算力最优的参数-数据比例，指出此前许多大模型训练不充分70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/688242API
curl https://kongchang.com/api/v1/knowledge/claims/688242MCP
get_claim(id=688242)