Unverified50% confidenceBenchmarkExact time
研究表明GPT-4作为评估者与人类专家评分的一致性超过80%，显著高于BLEU、ROUGE等基于词汇匹配的自动化指标
1
Sources
50%
Confidence
Long-term
Relevance
7/25/2026
First Seen
Sources
Related Claims
UnverifiedWhen a GPT-4-level model acts as a Judge, its evaluation results agree with human experts more than 80% of the time.84% similarUnverifiedIn standardized tests, GPT-4's performance has surpassed that of more than 90% of human test-takers.78% similarUnverifiedOptimized prompts can improve GPT-4's accuracy on specific tasks from 60% to above 95%.76% similarVerifiedModels such as GPT-4, Claude 3.5, and DeepSeek have achieved pass rates exceeding 80% on the HumanEval programming benchmark.74% similarUnverifiedLarge models such as GPT-4 and Claude have surpassed 85% Text-to-SQL accuracy on benchmarks such as Spider.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/613444API
curl https://kongchang.com/api/v1/knowledge/claims/613444MCP
get_claim(id=613444)