Unverified50% confidenceFactExact time
In the SilCC dataset, as naming difficulty increased, the effectiveness of all participants (humans and MLLMs) declined.
1
Sources
50%
Confidence
Long-term
Relevance
10/21/2026
First Seen
Sources
Related Entities
Related Claims
Unverified基准污染（Benchmark Contamination）导致HumanEval、MMLU等主流基准区分度下降，部分模型HumanEval得分超过95%、MMLU超过90%但实际能力平庸69% similarUnverifiedMultiple academic studies have shown that models scoring highest on mainstream benchmarks such as MMLU and HumanEval often perform significantly worse on novel questions of equivalent difficulty.67% similarUnverifiedIn the WOVEN evaluation, even the strongest MLLM falls far below human level, and failure modes recur across different model families, so simply scaling up model size cannot eliminate these deficiencies.67% similarUnverified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65% similarUnverifiedFor some diseases, recognition accuracy even decreased after synthetic oversampling.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1079455API
curl https://kongchang.com/api/v1/knowledge/claims/1079455MCP
get_claim(id=1079455)