中文

科学大型语言模型综述:从数据基础到智能体前沿

计算与语言 2025-10-21 v2 人工智能

摘要

科学大型语言模型(Sci-LLM)正在改变科学研究中知识表示、集成和应用的方式,但其进展受到科学数据复杂性的深刻影响。本综述提出了一个全面的、以数据为中心的综合分析,将 Sci-LLM 的发展重新构建为模型与其底层数据基底之间的共同演化。我们制定了科学数据的统一分类法和科学知识的分层模型,强调了将科学语料库与一般自然语言处理数据集区分开来的多模态、跨尺度和特定领域的挑战。我们系统回顾了近期的 Sci-LLM,从通用基础模型到跨越不同科学学科的专用模型,并对 270 多个预训练/后训练数据集进行了广泛分析,说明了为何 Sci-LLM 提出了独特的要求——异构、多尺度、充满不确定性的语料库,这些语料库要求保留领域不变性并实现跨模态推理的表示。在评估方面,我们考察了 190 多个基准数据集,并追踪了从静态考试向具有先进评估协议的面向过程和发现的评估的转变。这些以数据为中心的分析突显了科学数据开发中的持续性问题,并讨论了涉及半自动化标注流水线和专家验证的新兴解决方案。最后,我们概述了向闭环系统的范式转变,在该系统中基于 Sci-LLM 的自主智能体主动实验、验证,并为一个活的、不断演进的知识库做出贡献。总体而言,这项工作为构建值得信赖的、不断演进的人工智能(AI)系统提供了路线图,这些系统将作为加速科学发现的真正合作伙伴发挥作用。

关键词

引用

@article{arxiv.2508.21148,
  title  = {A Survey of Scientific Large Language Models: From Data Foundations to Agent Frontiers},
  author = {Ming Hu and Chenglong Ma and Wei Li and Wanghan Xu and Jiamin Wu and Jucheng Hu and Tianbin Li and Guohang Zhuang and Jiaqi Liu and Yingzhou Lu and Ying Chen and Chaoyang Zhang and Cheng Tan and Jie Ying and Guocheng Wu and Shujian Gao and Pengcheng Chen and Jiashi Lin and Haitao Wu and Lulu Chen and Fengxiang Wang and Yuanyuan Zhang and Xiangyu Zhao and Feilong Tang and Encheng Su and Junzhi Ning and Xinyao Liu and Ye Du and Changkai Ji and Pengfei Jiang and Cheng Tang and Ziyan Huang and Jiyao Liu and Jiaqi Wei and Yuejin Yang and Xiang Zhang and Guangshuai Wang and Yue Yang and Huihui Xu and Ziyang Chen and Yizhou Wang and Chen Tang and Jianyu Wu and Yuchen Ren and Siyuan Yan and Zhonghua Wang and Zhongxing Xu and Shiyan Su and Shangquan Sun and Runkai Zhao and Zhisheng Zhang and Dingkang Yang and Jinjie Wei and Jiaqi Wang and Jiahao Xu and Jiangtao Yan and Wenhao Tang and Hongze Zhu and Yu Liu and Fudi Wang and Yiqing Shen and Yuanfeng Ji and Yanzhou Su and Tong Xie and Hongming Shan and Chun-Mei Feng and Zhi Hou and Diping Song and Lihao Liu and Yanyan Huang and Lequan Yu and Bin Fu and Shujun Wang and Xiaomeng Li and Xiaowei Hu and Yun Gu and Ben Fei and Benyou Wang and Yuewen Cao and Minjie Shen and Jie Xu and Haodong Duan and Fang Yan and Hongxia Hao and Jielan Li and Jiajun Du and Yanbo Wang and Imran Razzak and Zhongying Deng and Chi Zhang and Lijun Wu and Conghui He and Zhaohui Lu and Jinhai Huang and Wenqi Shao and Yihao Liu and Siqi Luo and Yi Xin and Xiaohong Liu and Fenghua Ling and Yuqiang Li and Aoran Wang and Siqi Sun and Qihao Zheng and Nanqing Dong and Tianfan Fu and Dongzhan Zhou and Yan Lu and Wenlong Zhang and Jin Ye and Jianfei Cai and Yirong Chen and Wanli Ouyang and Yu Qiao and Zongyuan Ge and Shixiang Tang and Junjun He and Chunfeng Song and Lei Bai and Bowen Zhou},
  journal= {arXiv preprint arXiv:2508.21148},
  year   = {2025}
}