In August 2025, the "Opinions of the State Council on Deepening the Implementation of the" Artificial Intelligence+"Action" was released, and "Artificial Intelligence+" science and technology were included in the key actions, involving accelerating the process of scientific discovery, driving innovation in technology research and development models, and improving efficiency. With the accelerated iteration of the universal big model capability, AI for Science (AI4S) is becoming the core track of global technological innovation, sparking a paradigm revolution related to the underlying logic of scientific research. The "Research Report on the Development of Artificial Intelligence Industry (2025)" released by the China Academy of Information and Communications Technology shows that from 2023 to the first half of 2025, the investment and financing activity in the vertical field of artificial intelligence+scientific research in China will continue to increase. In the first half of 2025, the investment scale of AI scientific research applications nationwide will be about 1 billion yuan. However, in sharp contrast to the booming industry, there is a real contradiction facing researchers: existing scientific multimodal models often rely on massive and difficult to obtain data, and the training process is like a "black box" that is difficult to reproduce and improve. The demand for data is high, the ecosystem urgently needs to be cultivated, and international competition is intensifying... Researchers are exploring the reconstruction of the AI4S development paradigm, committed to driving "AI scientists" to independently make scientific discoveries. Is it really necessary to have billions of data to teach AI to understand science in order to implement AI4S with small data and open source? This is a question that lingers in the minds of many researchers, especially those in the AI4S field. At the beginning of 2026, a latest "Scientific Intelligence Practical Guide" will be released in the open source community. Research teams from Shanghai Jiaotong University, DP Technology, MemTensor, Institute of Theoretical Physics of the Chinese Academy of Sciences and other institutions jointly released the Innovator-VL multi-mode large model. Researchers say that the training process of Innovator VL has proven that without blindly piling up data, it can surpass many models with data volumes of billions in multiple scientific benchmark tests with less than 5 million carefully selected scientific training samples and transparent training strategies, breaking the "violent aesthetics" of the "data only theory" of scientific big models and verifying the technological path of "quality over quantity". The first author of the Innovator VL paper, Wen Zichen, told reporters, "In the absence of large-scale data, efficient and reproducible scientific multimodal models are not only possible, but also a practical path to future scientific discoveries." As a customized multimodal large model for the scientific field, Innovator VL has achieved cross scale and interdisciplinary full scenario scientific understanding capabilities. From molecular formulas, crystal structures, cryo electron microscopy images in the microscopic world, to astronomical light curves and remote sensing images in the macroscopic universe, to complex formulas and algorithm flowcharts in the field of mathematical logic, models can achieve deep analysis and logical reasoning. In practical cases, faced with the astronomical core task of analyzing celestial light variation curves, rigorous logical deduction was completed through the characteristics of light variation curves and the evolution laws of bands, and the model accurately identified the light variation characteristics of Type Ia supernovae; In organic chemistry scenarios, models identify reactant structures, decompose functional group characteristics, match reaction types and options, and assist in chemical reasoning. It is worth noting that, unlike the conventional open source model that only opens model weights in the industry, the Innovator VL R&D team has fully open sourced an end-to-end reproducible full process development pipeline, covering data collection and cleaning methodology, complete instruction fine-tuning and reinforcement learning strategies, hyperparameter optimization schemes, and evaluation frameworks. Assistant Professor Zhang Linfeng from the School of Artificial Intelligence at Shanghai Jiao Tong University believes that the role of AI in scientific research is gradually evolving from an "acceleration tool" to a "cognitive participant". In the past, we focused more on making the model process data faster, but in the future, the more critical issue is whether it can participate in the definition and reconstruction of scientific problems themselves, "said Zhang Linfeng. From this perspective, the significance of Innovator VL lies not only in performance improvement, but also in a forward-looking exploration of whether AI can become a part of the scientific research process. This exploration may redefine the boundaries of human-machine collaboration in scientific discovery. The AI4S field, represented by the Innovator series models, has made continuous breakthroughs and fully opened up, providing efficient and reusable research paradigms for researchers, and enabling universities, small and medium-sized research institutions lacking massive computing power and data resources to participate in scientific and intelligent innovation exploration at low cost. E Weinan, an academician of the CAS Member and chief consultant of the Artificial Intelligence School of Shanghai Jiaotong University, said frankly that the key infrastructure of AI4S has gradually taken shape, and the new era of Agricultural Science at Scale (large-scale autonomous agent research) has officially opened. The success of the Innovator series model, which breaks through the entire chain of China's AI4S racing new track, is just a microcosm of the vigorous development of China's AI4S. Currently, China is forming a full chain development pattern of "breakthrough in the base model - transfer of scientific research capabilities - landing of industrial scenarios", assisting researchers in solving the core pain points of traditional scientific research, such as long cycles, high costs, and difficult trial and error, and striving to compete for the global AI4S research highland. Guo Yike, a foreign academician of the Chinese Academy of Engineering and Chief Vice President of the Hong Kong University of Science and Technology, believes that AI4S is not only a technological revolution, but also a cognitive revolution, which is promoting the transformation of research paradigms from "trial and error driven" to "data+model driven", upgrading AI from a passive efficiency tool to a research partner that can actively reason and evolve independently. The characteristic path of "industry demand driven+industry university research collaborative innovation" has formed a positive cycle of basic innovation and industrial landing, and open source achievements have greatly reduced the R&D threshold of the industry. However, the large-scale implementation of AI4S still faces many common challenges, including multimodal information alignment, scarcity of high-quality standardized data, model illusion and verifiability bottlenecks, as well as interdisciplinary talent gaps, lack of industry standards, and high computing power costs. In the future, with the continuous deepening of industry university research collaboration and the continuous improvement of the open source ecosystem, China's AI4S innovation will continue to contribute reusable Chinese solutions to the global paradigm shift in scientific research, promoting AI to truly become the "best assistant" and "super partner" for researchers. (New Society)