Unverified50% confidenceBenchmarkExact time
一项针对GPT-4的研究显示，对MMLU题目进行语义等价改写后，模型准确率在部分子集上下降超过10个百分点
1
Sources
50%
Confidence
Long-term
Relevance
7/5/2026
First Seen
Sources
Related Claims
UnverifiedA 2023 Stanford University study showed that multimodal models such as GPT-4V, when parsing flowcharts with nested conditional logic, had accuracy approximately 15–25 percentage points lower than equivalent textual descriptions.79% similarUnverifiedA 2023 Stanford University research experiment demonstrated that when key information was placed in the middle of an input document, GPT-4's correct recall rate dropped approximately 20–30% compared to when the information was placed at the beginning or end.75% similarUnverifiedCompared to GPT-5.2, GPT-5.4 reduces the probability of an individual statement being incorrect by 33% and the probability of an entire response containing any error by 18%.74% similarUnverifiedDuring large model inference, actual GPU utilization is often less than 10%.73% similarUnverifiedSome models have already surpassed 90% of the human average score on MMLU, causing their discriminative power to drop sharply.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/112194API
curl https://kongchang.com/api/v1/knowledge/claims/112194MCP
get_claim(id=112194)