BEIJING -- A joint research team has developed a general artificial intelligence (AGI) system capable of both autonomous problem proposing and automated problem solving, marking a critical milestone in the self-developed logic cores for automated reasoning.

In performance and functional diversity, the system, TongGeometry, has fully outperformed international benchmarks, including DeepMind's AlphaGeometry. This represents a major step forward in AI-assisted mathematical research and the localization of intelligent education.

The study, jointly conducted by the Beijing Institute for General Artificial Intelligence (BIGAI), the School of Psychological and Cognitive Sciences at Peking University, the School of Intelligence Science and Technology at Peking University, the Institute for Artificial Intelligence at Peking University, and the Wuhan Institute for Artificial Intelligence at Peking University, was published in the journal Nature Machine Intelligence on Monday.

Mathematics Olympiads have long served as the litmus test for AI's logical reasoning capabilities. In early 2024, DeepMind's AlphaGeometry made global headlines by showcasing AI's enormous potential for problem-solving. However, AlphaGeometry is essentially a "passive solver" whose training relies heavily on large-scale synthetic datasets and costly computational resources.

In contrast, the independently developed TongGeometry exhibits a higher dimension of intelligence. It is not merely an "honor student" capable of scoring full marks, but also a "master teacher" capable of creating elegant and novel mathematical problems.

"We identified a profound duality in our research: when the proof difficulty of a geometric proposition is far higher than its construction complexity, it possesses 'aesthetic value' as an Olympiad-level problem," said Zhang Chi, the first author of the paper and a researcher at BIGAI.

"By modeling this duality, TongGeometry can precisely capture high-quality problems that meet the aesthetic standards of human mathematicians from a vast pool of spatial combinations. This is a global first, representing a paradigm shift from 'imitative solving' to 'autonomous creation'," Zhang noted.

TongGeometry clearly highlights the superiority of original domestic technology in terms of performance. While AlphaGeometry requires massive computing clusters, TongGeometry can solve all International Mathematical Olympiad geometry problems from 2000 onward in 38 minutes or less using just a single consumer-grade GPU.

Its reasoning efficiency and accuracy have reached world-leading levels. Furthermore, the system utilizes an innovative normalized representation technology to compress the search space by several orders of magnitude, effectively solving the path explosion problem inherent in traditional methods.

"The significance of TongGeometry lies not only in the increase in solving speed but in its realization of the 'small data, big task' paradigm by simulating the intuition and aesthetics of human mathematicians," said Zhu Yixin, assistant professor from the School of Psychological and Cognitive Sciences at Peking University.

"This path, which does not depend on massive labeled data but evolves through internal logic, is the key to the development of AGI. Our system not only benchmarks against the most advanced international AI but also leads the way in understanding the underlying aesthetics of logic and the autonomous discovery of scientific laws," Zhu said.

Three new geometry problems autonomously generated by the system were officially selected for the 2024 Chinese Mathematical Olympiad (Beijing District).

This breakthrough provides core technical support for future advances in automated mathematical proofs, personalized intelligent education, and the development of "Science Large Language Models."

Going forward, the joint research team will continue to iterate on the "Tong" series of general intelligence models, driving Chinese AI technology to take the lead in more fields of complex logic and scientific discovery.