At the World Artificial Intelligence Conference held not long ago, the “Rock and Science Foundation Big Model 2.0”, jointly developed by the Chinese Academy of Sciences across multiple disciplines, was officially unveiled. It surpassed the international top general-purpose big models in more than 60 specialized scientific research missions, and also ran at second-rate speeds in astronomy, mechanics, chemistry, and other fields, drawing attention from the industry. The reporter from the Global Times recently interviewed Li Linjing, a core member of the model research and development team and a researcher at the Chinese Academy of Sciences‘ Automation Research Institute, to deeply disassemble the core breakthroughs, practical value, and industry advantages of Rock 2.0.
All-in-One: Making the Scientific Brain No longer “patch”
“You can imagine Rock 1.0 as a ‘project manager’. When it encounters a professional problem, it calls for expert models in various fields to help. It‘s a relatively hard ‘general-specialty combination’.” Li Linjing said that Rock 2.0‘s most core breakthrough was to “natively” train general and professional abilities together in the same foundation model, truly achieving general-specialty integration.
What does such integration mean for frontline scientific research? Li Linjing says: “The most intuitive value is that scientific researchers no longer need to learn and master various complex professional models and tools. Through this one foundation model, they can complete the entire process from data understanding, reasoning, prediction, to specialized content generation.” The deeper value is still behind. In the past, under the hard combination method, professional models such as mathematics, physics, and chemistry separated each other‘s abilities, but after integration, the various professional abilities internalized into the foundation model can learn from each other and influence each other. “For example, physical intuition in material design may suddenly inspire chemical reaction pathways, making interdisciplinary original thinking more likely to collide.”
Behind this breakthrough is a progressive three-tier architecture built by the R&D team: unified coding of scientific data, alignment of knowledge of the natural world, and domain task-oriented decoding. Based on 8 million high-quality scientific inference data gathered from the Chinese Academy of Sciences‘ scientific and technological corpus, the team transformed scientists‘ way of thinking into inference guidelines, making model learning “highly reliable, strongly logical, and explainable” based on inference based on chains of knowledge and evidence. Li Linjing revealed that in the assessment of more than 60 specialized scientific research tasks, Rock 2.0 significantly surpassed flagship models such as GPT-5.5 and achieved international best results in tasks such as predicting chemical properties and predicting protein loci. A scientific mind that is both erudite and specialized is freeing researchers from fragmented tools.
“Efficiency increased by 50 times compared to expert visual inspection”
In order to make the journalist truly feel the leap in efficiency, Li Linjing cited an example from the astronomy field. In the past, rare celestial body candidate review was a delicate task that relied heavily on expert experience. Experts needed to carefully examine with the naked eye key features in the spectrum, determining whether it was contaminated and whether the signal was reliable. Processing a spectrum often took about 10 seconds.
Rock 2.0‘s approach is completely different. “We let the model learn the expert‘s censorship logic: first observe the whole, then magnify the local, actively invoke spectral visualization tools, gradually examine key bands, eliminate contamination and obtain local evidence, ultimately forming explainable, traceable censorship conclusions.” Li Linjing says that this way, the agent can process a spectrum on average in about 0.2 seconds, bringing about a 50-fold increase in efficiency compared to expert artificial visual inspection. It is understood that this set of astronomical spectral agents has been deployed at the National Observatory, directly accelerating the entire process from observation to analysis by the Guo Shoujing Telescope, increasing the accuracy of rare celestial bodies identification by about 50% compared to the existing optimal method, and compressing the parameter simulation duration from several hours to minutes.
Efficient agents are not only used to look up at the starry sky. In the mechanical case shown at the conference, based on Rock‘s ability to understand and infer field data, fluid simulation software implemented a new paradigm of “gridless, second-rate response, high throughput evaluation,” producing simulation results within 10 seconds with error control within 5%. The chemistry field is even more exciting. The rock-based machine scientist system, which autonomously orchestrates the co-operation of 5 agents such as literature mining, experimental design, and machine operation, has completed 462 sets of sample synthesis characterization and tens of millions of levels of virtual screening. Eight of the top 20 candidate materials perform better than the original optimal samples, and have been running steadily for more than 1,500 hours. In addition, the literature compass agent can generate professional reviews within 3 hours, with an accuracy rate of evidence attribution reaching 90%, accumulating nearly 40,000 reviews. Scientific research is moving from “people looking for tools” to “agent autonomous planning”.
National computing power builds foundations, heading to the “next crossroads” of self-discovery
Currently, AI4S (Artificial Intelligence Enablement Scientific Research) has become the core new track for global technology competition, and countries are all increasing their layout. In Li Linjing‘s view, our country‘s AI empowerment basic scientific research is currently still in the preliminary exploration stage. Currently, most scientific research teams in China mainly deploy general large models and specialized tools to conduct research. Vertical specialized models that are fine-tuned by various institutions are limited by data and parameters, and have limited support for real scientific research. Meanwhile, the establishment of the rock model system has paved the way for China‘s AI4S track to develop a distinctive path of institutionalization synergy, general specialty integration, and industrial and scientific research fusion.
According to the Huanqiu Times reporter, relying on the multidisciplinary institutionalization advantages of the Chinese Academy of Sciences, the Rock Project has built an innovation union across fields, disciplines, and systems, gathering dozens of research institutes, universities, and industry leaders within the institute to collaborate in the joint effort. Among them, the automation research institute leads the core technology joint effort, related companies take over industrialization and tool base construction, several specialized research institutes deepen the vertical field model research and development, and several domestic companies provide computing power support, forming a comprehensive, three-dimensional, joint effort matrix. Currently, the Rock System has landed in more than 50 Chinese Academy of Sciences research institutes, more than 30 external universities, and central enterprises, and is reaching the world through international cooperation platforms.
Comparing the world‘s top AI4S technology, Li Linjing clearly outlined our country‘s differentiation advantages and iterative directions. “Our country‘s big scientific model‘s biggest characteristics are strong knowledgeability and broad disciplinary coverage. Relying on the nation‘s vast scientific research resources and scientists collaborating to cross the hurdles, we have built a scientific material system that covers multiple fields, able to accurately adapt to various scientific research scenarios, and deeply understand specialized scientific data.” He also frankly stated that the advantages of America‘s top models lie in deep cultivation of single-point technology, with higher accuracy in segmental fields such as pharmaceutical and single-material prediction, which is a direction that needs to be continuously pursued in the future.
But in the long run, our country has irreplaceable resource advantages. Li Linjing believes, “We have the world‘s largest number of large scientific devices, and a large number of high-performance devices are in the planning process of construction. These high-quality scientific data resources will continue to solidify the foundation for the development of our country‘s large scientific models, becoming the core foundation for curve overtaking.”
Facing the industry‘s development in the next 3-5 years, Li Linjing made a clear prediction: AI-powered research will completely abandon the “tool-assisted” model and move toward a new paradigm of autonomous research. “In the future, AI will no longer be just a tool to assist research. It will gradually possess the ability to autonomously dismantle research tasks, design experimental schemes, and complete verification iterations. Human scientists will take on more roles of supervision, decision-making, and innovation leadership.” He said that there are currently worldwide cases of AI assisting original research breakthroughs, and within the next 3 years, such results will become the mainstream of the industry. “In the direction of scientific research that does not require wet experiments and can rely on autonomous laboratories, the AI autonomous research paradigm will be fully grounded. The Rock team will also continue to deepen the integrated technology route, expand multidisciplinary application scenarios, and continue to perfect autonomous and controllable scientific intelligence systems. With AI as a wing, we will continue to unlock the infinite possibilities of basic scientific research, helping our country‘s cutting-edge scientific research achieve more original and important breakthroughs.”