研究团队介绍说，磐石·科学文献解析器从底层算法出发，通过构建面向科学语义理解的多模态训 练体系与强化学习机制，在公式、文本、图表等多元素协同解析上实现质的飞跃，为全球科研工作者提 供真正"懂科学"的智能解析引擎。 在磐石·科学文献解析器研发过程中，团队摒弃仅依赖通用视觉语言大模型的思路，转而构建一套 专为科学文献场景量身定制的算法训练范式。其核心在于三大技术支柱：全场景覆盖的科学数据构建、 多模态监督微调策略，以及面向科学文献语义的强化学习优化机制。 在数据层面，系统性采集并构建覆盖手写体、数字排版体与纸质扫描体三大典型科学书写形态的训 练语料，这一"全形态、多学科、高质量"的数据基础，为模型理解科学表达的复杂性提供了坚实支撑。 模型训练阶段采用两阶段优化策略：首先通过多模态有监督微调，使模型初步掌握文本、公式、表 格、插图等异构元素的联合表征能力。在此基础上，引入一种面向科学文献语义的梯度强化学习策略优 化框架，实现模型不仅"看得清"，更能"理解对"。 记者11月1日从中国科学院自动化研究所（自动化所）获悉，该所"AI+科学"研究团队近日正式推出 新一代科学文献解析工具——磐石·科学文献解析器，为全球科研工 ...

"为了解决这些难题，中国科学院系统内多家研究所组成研发团队，开展体系化研究，推动'AI+科 学'研究向平台化、体系化转变，成功打造了磐石·科学基础大模型。"中国科学院科技基础能力局副局 长、中国科学院自动化研究所副所长曾大军介绍，该模型就像一个跨学科的"智能操作系统"，可统一管 理数据和模型等资源，调度计算仿真等各类工具，深度赋能"假设提出—方案规划—仿真推演—实验验 证—规律发现"的科研全流程。 "AI+科学"成为新趋势，为解决重大科技难题带来了前所未有的机遇。7月26日，在2025世界人工 智能大会上，中国科学院正式发布磐石·科学基础大模型。该模型是一个专门为科研打造的智能平台， 由专业科学知识和数据训练而成，能深入理解波、谱、场等各种科学数据，并具备深度文献分析、科学 知识推理和科研工具编排能力。 "AI+科学"研究常用两种方式：用专业数据微调通用模型，或自建单一领域专用工具。这带来了科 学数据不互通、专业推理能力不够以及研发生态封闭三大难题。 "目前，磐石·科学基础大模型已经在多个学科领域进行了深入应用，大幅加速了科研进程。"曾大 军举例说，在高能物理领域，北京正负电子对撞机的研究人员依托磐石·科学基 ...

Core Viewpoint - In 2025, large models are transitioning from laboratories to practical applications across various industries, showcasing significant breakthroughs in multiple scientific fields, including quantum computing, life sciences, materials science, earth sciences, and deep space astronomy [1][2]. Breakthrough Achievements - The world's first AI-based quantum computing neutral atom arrangement algorithm was developed, capable of arranging 2024 quantum bits in 60 milliseconds, overcoming traditional time constraints associated with increasing atom numbers [5]. - The first multi-agent virtual disease scientist system, "OriGene," was introduced, which can automatically discover and validate new treatment targets for cancers, establishing a complete intelligent process from data to verification [6]. - The first single-cell DNA methylation model, scDNAm-GPT, was launched, achieving over 90% accuracy in early detection of various cancers and respiratory diseases using blood samples [7]. - The "Fengdeng" breeding model was established, marking a breakthrough in China's seed industry technology, with over 100 breeding units already testing its applications [8]. - An intelligent scientific discovery system for condensed matter science was released, achieving practical-level standards for new copper-based superconducting materials [9]. - The EarthLink AI Earth scientist system was developed, significantly enhancing research efficiency by 160 times through automated experimental design and data analysis [10]. - An AI tracking system for space debris was created, improving tracking accuracy by 70% compared to traditional methods [11]. - The ChemBOMAS multi-agent system optimized chemical reactions, reducing precious metal catalyst usage by 90% and increasing yield from 20% to 96% [12]. - The RNA virus language model, Viracle, was introduced, providing high-precision predictions for RNA viruses, with over 95% accuracy for human viruses [13]. - A 3D aircraft generation agent was developed, streamlining the aircraft design process through natural language interaction and rapid concept generation [14]. Future Directions - The Shanghai AI Laboratory plans to continue advancing the integration of specialized technologies to address key scientific discovery challenges, leveraging the "Shusheng" scientific discovery platform for innovation [15].