SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines
计算与语言
2025-12-16 v3
摘要
我们提出一种科学推理基础模型，将自然语言与异构科学表征相 ALIGN。该模型在覆盖 2060 亿 token 的科学语料库（包括科学文本、纯序列和序列-文本对）上进行预训练，随后通过 4000 万指令进行 SFT，对 cold-start bootstrapping 进行退火式处理以激发长链推理，并通过基于任务特定奖励的强化学习，使其具备 deliberate scientific reasoning 能力。它支持四类能力，覆盖最多 103 项任务：(i) 在文本和科学格式之间进行忠实翻译，(ii) 文本/知识抽取，(iii) 属性预测，(iv) 属性分类，(v) 无条件和条件序列生成与设计。与专门系统相比，我们的方法扩大了指令覆盖范围，提高了跨域泛化能力，并增强了忠实度。我们详细阐述了数据策划和训练过程，展示跨学科学习加强了迁移能力和下游可靠性。该模型、指令调优数据集以及评估代码已开源于 https://huggingface.co/SciReason 和 https://github.com/open-sciencelab/SciReason。
引用
@article{arxiv.2509.21320,
title = {SciReasoner: Laying the Scientific Reasoning Ground Across Disciplines},
author = {Yizhou Wang and Chen Tang and Han Deng and Jiabei Xiao and Jiaqi Liu and Jianyu Wu and Jun Yao and Pengze Li and Encheng Su and Lintao Wang and Guohang Zhuang and Yuchen Ren and Ben Fei and Ming Hu and Xin Chen and Dongzhan Zhou and Junjun He and Xiangyu Yue and Zhenfei Yin and Jiamin Wu and Qihao Zheng and Yuhao Zhou and Huihui Xu and Chenglong Ma and Yan Lu and Wenlong Zhang and Chunfeng Song and Philip Torr and Shixiang Tang and Xinzhu Ma and Wanli Ouyang and Lei Bai},
journal= {arXiv preprint arXiv:2509.21320},
year = {2025}
}
备注
technical report