Innovative research paradigms, AI models, and autonomous robots at the 2025 world artificial intelligence conference
The 2025 World Artificial Intelligence Conference (WAIC) and High-Level Meeting on Global AI Governance was held in Shanghai from July 26–29, 2025. During this conference, many new artificial intelligence (AI) ideas were shared by leading AI pioneers, such as Geoffrey Hinton, Yoshua Bengio, and Andrew Chi-Chih Yao. More than 40 new AI models were released, and more than 60 robots were demonstrated.
Innovative research paradigms of AI for science
AI for science dominated the agenda. In the AI for Life Science Forum, academician Jingsong Li from the Shanghai Institute of Biochemistry and Cell Biology, Chinese Academy of Sciences, referenced a review article1 published in The Innovation Life in 2024 to show the critical role of high-quality big data as the foundational resource for AI-driven biomedical research. Additionally, in another session, academician Weinan E highlighted the evolving landscape from model-centric to data-centric AI development. Despite the proliferation of approximately 3,755 large language models (LLMs) globally and 1,509 in China, most are pre-trained on similar natural language corpora sourced from the Internet. As noted by Li Xu, Chief Executive Officer of SenseTime, the extensive extraction of natural language data from online sources is approaching saturation, prompting a shift toward underexplored data domains such as biomedical datasets, which are substantially larger and more complex.
The challenges associated with biomedical big data are elaborated in a perspective2 published in The New England Journal of Medicine in 2025. Despite the abundance of real-world data (RWD), there is a scarcity of actionable real-world evidence. The inherent difficulties in biomedical data mining persist,3 including the enormous data volume, heterogeneity of data types, continuous data generation, and low data value density.4 Although multimodal LLMs have been developed, their capacity to comprehensively process diverse RWD remains limited.
Next-generation AI models and intelligent agents
Various AI models and agents tailored for different types of RWD and applications should be integrated within the AI operating system (OS), such as the Bio-medical Big Data Operating System (Bio-OS; https://bio-os.github.io/) and Biomni, a general-purpose biomedical AI agent (https://biomni.stanford.edu/). During the 2025 WAIC, the Chinese Academy of Sciences (CAS) launched ScienceOne (https://scienceone.ia.ac.cn/), which encompasses S1-Literature, with 170 million scientific publications and real-time open-source scientific data, along with S1-ToolChain, featuring over 300 research tools spanning digital cell analysis, microphysics, and crystal engineering. Additional notable AI models introduced at the 2025 WAIC are summarized in Table 1. For instance, Intern-S1 (https://chat.intern-ai.org.cn/), developed by Shanghai AI Laboratory, is an advanced multimodal reasoning platform that combines robust general-task capabilities with state-of-the-art performance across scientific domains, such as chemical structure interpretation, protein sequence analysis, and synthesis pathway planning, positioning it as a versatile scientific research assistant. OriGene (https://github.com/GENTEL-lab/OriGene), from Lin Gang Laboratory, is an open-source, self-evolving, multi-agent system functioning as a virtual disease biologist to identify and prioritize therapeutic targets at scale. SeedLLM (https://seedllm.org.cn/), from Yazhouwan National Laboratory, integrates an LLM with the Rice biological knowledge graph (RBKG), encompassing genomic annotations and transcriptomic and proteomic data from over 1,800 studies. EarthLink (https://earthlink.intern-ai.org.cn), developed by the Institute of Atmospheric Physics, CAS, is a self-evolving AI agent dedicated to climate science, automating comprehensive research workflows from planning and code generation to multi-scenario analysis, delivering scientifically rigorous insights.
