论文40
Epoch AI:模型自评偏好自身输出
原标题:We also tested whether models could recognize how they fell short when given gold-standard human output and our full grading rubric. They...
我们还测试了模型在获得金标准人类输出和完整评分标准后,能否认识到自己的不足。它们不认同人类的评估,除一种情况外,其余均偏好自己的输出。 模型常常夸大人类输出中的微小瑕疵,并优先考虑表面因素,例如所运行实验的数量。
中文译文尚未完成,先为你展示原文。
We also tested whether models could recognize how they fell short when given gold-standard human output and our full grading rubric. They disagreed with human assessments, preferring their own outputs in all but one case.
Often models exaggerated trivial flaws in the human output and prioritized superficial factors, such as the number of experiments run.
