Unverified50% confidenceBenchmarkExact time
Even when the correct value was retrieved, agents acted on it only 55% of the time, exposing a gap between retrieving information and acting on it.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMedIQ多智能体系统最终达到57%准确率，摘要智能体带来约5%提升，而拥有全部前置信息的Oracle上限为66%72% similarUnverifiedIn the ItaSoRL experiment, the accuracy of probing the agent's internal states was only approximately 50%, equivalent to random guessing.69% similarUnverifiedSingle-period data, lack of horizontal comparison, and differences in reporting standards mean that the 51% zero-payout data should be treated as a signal worth further investigation rather than a definitive judgment of the company.67% similarUnverified研究表明人类评估者在判断LLM输出的事实准确性时，正确率往往只有50%-70%67% similarUnverified当综合评估指数中顶尖模型得分接近天花板时会出现基准饱和问题，60分对59分的差距在统计学上通常处于误差范围之内66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/892959API
curl https://kongchang.com/api/v1/knowledge/claims/892959MCP
get_claim(id=892959)