Unverified50% confidenceBenchmarkExact time
The Gorilla benchmark released by the University of California, Berkeley shows that when the number of candidate tools exceeds 20, even models at the GPT-4 level experience a 15%-30% drop in tool selection accuracy.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/6/2026
First Seen
Valid until: 12/5/2026
Sources
Related Entities
Related Claims
UnverifiedIn vertical domains such as industrial defect detection and pathology slide analysis, the zero-shot detection accuracy of large models such as GPT-4V and Gemini lags behind specialized models by 20–40 percentage points on the mAP metric.68% similarUnverified在工具数量超过20个的场景下，工具描述的语义歧义度与模型选择准确率之间存在显著的负相关关系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% similarVerifiedWhen available tools exceed 15-20, mainstream LLMs show a noticeable decline in tool selection accuracy.65% similarUnverifiedMultiple benchmark tests in 2023 (such as the HaHackathon dataset) showed that top LLMs still have significantly lower accuracy than ordinary human annotators on sarcasm recognition tasks64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/865049API
curl https://kongchang.com/api/v1/knowledge/claims/865049MCP
get_claim(id=865049)