FirstPass: Grounding AI Scientific Judgment in Multi-Round Editorial Outcomes
FirstPass: 在多轮编辑结果中奠定AI科学判断的基础
机构 * The University of Texas at Austin(德克萨斯大学奥斯汀分校) ; RediMinds Inc.(RediMinds公司) ; Disaster Science Operations Center, Western Kentucky University(西肯塔基大学灾害科学运营中心)
AI总结 针对同行评审AI在领域覆盖、对话建模和评估标准上的不足,提出FirstPass数据集和微调模型,利用Nature Communications多轮评审对话,通过响应损失掩码实现80.5%的编辑结果预测准确率,并生成接近人类长度的评审意见。
Comments Accepted at the AI for Science Workshop at the 43rd International Conference on Machine Learning (ICML 2026). 9 pages, 2 figures, 6 tables