BEIJING, Jan. 28 (Xinhua) — A joint research team has developed an artificial general intelligence (AGI) system capable of both autonomous problem proposal and automated problem solving, marking an important milestone in the homegrown logic core for automatic reasoning.
In terms of performance and feature diversity, the system TongGeometry completely outperforms international benchmarks, including DeepMind’s AlphaGeometry. This represents a major step forward in AI-assisted localization of mathematics research and intelligent education.
The study was jointly conducted by Beijing Institute of Artificial Intelligence (BIGAI), Peking University School of Psychological and Cognitive Sciences, Peking University Academy of Intelligence Science and Technology, Peking University Institute of Artificial Intelligence, and Peking University Wuhan Institute of Artificial Intelligence, and was published Monday in the journal Nature Machine Intelligence.
Math Olympiad has long served as a litmus test for AI’s logical reasoning abilities. In early 2024, DeepMind’s AlphaGeometry made global headlines by demonstrating the huge potential of AI in problem-solving. However, AlphaGeometry is essentially a “passive solver” and its training relies heavily on large synthetic datasets and expensive computational resources.
In contrast, the proprietary TongGeometry exhibits a higher level of intelligence. He is not just an “honor student” who can get perfect scores, but also a “master teacher” who creates elegant and innovative math problems.
“We identified a deep duality in our study: when the difficulty of proving a geometric proposition is much higher than the complexity of its construction, it has ‘aesthetic value’ as an Olympic-level problem,” said Zhang Chi, first author of the paper and a researcher at BIGAI.
“By modeling this duality, TongGeometry is able to accurately capture high-quality problems that meet the aesthetic standards of human mathematicians from a vast pool of spatial combinations. This is a world first and represents a paradigm shift from ‘mimetic solution’ to ‘autonomous creation’,” said Zhang.
TongGeometry clearly shows the superiority of domestically produced technology in terms of performance. While AlphaGeometry requires a large compute cluster, TongGeometry can solve every International Mathematics Olympiad geometry problem since 2000 in under 38 minutes using just one consumer-grade GPU.
Its inference efficiency and accuracy have reached the world’s top level. Additionally, the system utilizes innovative regularized expression technology to compress the search space by several orders of magnitude, effectively solving the path explosion problem inherent in traditional methods.
“The importance of TongGeometry lies not only in improving the solution speed, but also in realizing the ‘small data, big tasks’ paradigm by simulating the intuition and aesthetics of human mathematicians,” said Zhu Yixin, assistant professor in the School of Psychological and Cognitive Sciences at Peking University.
“This path of evolving through internal logic, without relying on large amounts of labeled data, is key to AGI development. Our system not only benchmarks itself against the state-of-the-art international AI, but also leads in understanding the aesthetics of the underlying logic and autonomous discovery of scientific laws,” Zhu said.
Three new geometry problems autonomously generated by the system have been officially selected for the 2024 China Mathematics Olympiad (Beijing Area).
This breakthrough provides core technical support for future advances in automated mathematical proofs, personalized and intelligent education, and the development of “scientific large-scale language models.”
In the future, the joint research team will continue to iterate the “Tong” series of general intelligence models, promoting China’s AI technology to take the lead in more areas of complex logic and scientific discovery. ■
