In 2025aaaiConference: China’s scientific research power rises, and large model technology innovation leads the future
2025From February 25th to March 4th, the annual AAAI AI Top Conference was held in Philadelphia, USA. A total of 12,957 valid submissions were received at this conference, and 3,032 were accepted, with an acceptance rate of 23.4%. It is worth noting that Chinese universities and enterprises have demonstrated strong scientific research strength at this conference, and many papers have been selected, many of which focus on breakthroughs in large model technology.
Papers from a number of Chinese institutions were selected, demonstrating their scientific research strength
A number of papers from Chinese universities and Internet companies were included in AAAI 2025. For example, the School of Information of Xiamen University has 32 papers selected, covering the fields of multimodality, object detection and autonomous driving; 14 papers were selected in the field of data science and analysis at the Hong Kong University of Science and Technology, including large language model compression, fine-tuning, supervised learning, and AI image processing.quick worker7 papers were selected, and the research directions involved short video processing, video Q&A technology and recommendation models; Ant Group has 18 papers selected, 3 of which are Oral papers, covering enhanced large model privacy protection, improving inference speed and ability, improving large model training efficiency, and reducing model hallucinations.
Large model technology optimization has become the focus
At this conference, many papers focused on the technical optimization and breakthrough of large models, reflecting research in this fieldhot spot。
The Oral paper “Cross-domain Fine-Tuning Algorithm for Large Models with Variable Privacy-Practicality” jointly completed by Ant Digital, Zhejiang University, University of Liverpool and East China Normal University proposes a cross-domain fine-tuning framework called ScaleOT. The framework provides a variety of lossy compression emulators at different scales for model privacy and supports lossless fine-tuning. The experimental results show that ScaleOT achieves a 12.5% increase in the achievement rate of full fine-tuning on a variety of models, while saving 90% of the computing power requirements, providing a feasible solution for lightweight cross-domain fine-tuning of tens of billions of parameter models. At present, this algorithm has been applied to Ant Digital’s Morse large model privacy protection products, and has become one of the first products in China to pass the special test of trusted execution environment products of the large model of the Academy of Information and Communications Technology.
Another Oral paper completed by the joint laboratory of Ant Group and Zhejiang University, “K-ON: Stacking Knowledge Graph Knowledge at the Output Layer of Large Models”, proposes a method called K-ON, which uses a multi-step token prediction mechanism to generate evaluation results for all entities, and realizes comparative learning at the graph entity level adapted to the language model. Experiments show that the performance of K-ON is better than the traditional method and some multimodal methods that utilize additional text and visual information, and the number of training rounds is reduced from 1000 rounds to 5 rounds, and the training can be completed in 1 hour with 8 A100 graphics cards on the DB15K dataset.
China’s scientific research forces continue to rise
In recent years, China’s influence at top global AI conferences has grown. According to the China Science and Technology Information Statistical Report 2024, China ranks first in the world in terms of the number of hot papers (48.4%), the second in the world in the number of highly cited papers (33.8%), and the number of citations in seven disciplines such as computer science. The Stanford AI Index 2024 also pointed out that the industry is currently leading the way in cutting-edge research in artificial intelligence, and Chinese Internet companies and technology companies have become a scientific research force that cannot be ignored. The strong alliance of industry, academia and research will help China achieve more achievements on the global frontier scientific and technological stage.
The above is the global AI summit AAAI 2025 For more details on the outstanding performance of the “Chinese team” jointly established in the United States, please pay attention to other related articles on the PHP Chinese website!
