Unverified50% confidenceBenchmarkExact time
在一个强化学习实验中，未设置识别虚假世界奖励时，数字生物判断自己是否处于虚拟世界的准确率仅约50%，接近随机猜测
1
Sources
50%
Confidence
Long-term
Relevance
8/9/2026
First Seen
Sources
Related Claims
UnverifiedAfter the researchers made the physical rules of the fake world affect the difficulty of food capture, after 10 rounds of testing, the digital creature's accuracy in identifying the fake world jumped from approximately 50% to approximately 73%.77% similarUnverifiedIn the ItaSoRL experiment, the accuracy of probing the agent's internal states was only approximately 50%, equivalent to random guessing.64% 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% similarUnverified在机器学习多种子实验中重复次数通常只有5到10次，样本量极小，正态性假设难以验证，因此 Wilcoxon 检验往往是更保守稳健的选择62% similarUnverified泰特洛克的研究得出结论：专家预测的准确率仅略高于随机猜测，甚至不如简单的统计基准线62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/717713API
curl https://kongchang.com/api/v1/knowledge/claims/717713MCP
get_claim(id=717713)