On November 28, Chinese artificial intelligence (AI) company DeepSeek announced the AI model DeepSeekMath-V2, which is considered a breakthrough in the field of AI mathematical reasoning, setting new performance standards and expanding the limits of problem-solving capabilities using machine learning.
DeepSeekMath-V2's source code is publicly available on Hugging Face and GitHub.
The model integrates a self-verification framework to check the validity of a chain of arguments, in addition to generating correct answers, something that many current AI models still struggle with.
The evaluation results show that DeepSeekMath-V2 is qualified for gold medals at the 2025 International Mathematical Olympiad (IMO) and the 2024 Chinese Mathematical Olympiad (CMO).
Notably, the model scored 118/120 points in the 2024 Putnam exam, far surpassing the 90-point mark, the highest record ever achieved by humans.
The model's logical reasoning capabilities are tested using the IMO-ProofBench system, a dedicated benchmark used to verify the reasoning accuracy of AI models.
DeepSeekMath-V2 demonstrates superior performance compared to many other state-of-the-art models, including DeepMind's DeepThink.
IMO-ProofBench operates on a cross-checking mechanism: one model takes on the role of “proving,” generating a chain of mathematical arguments, while the other model takes on the role of “validating,” assessing the strength of the argument.
This mechanism allows for the detection of errors in the model's thinking, an inherent weakness of contemporary AI systems.
According to the development team, DeepSeekMath-V2's self-verifying method helps solve the biggest limitation of current AI models: the ability to generate correct answers but based on incorrect or inconsistent reasoning.
DeepSeek believes that these advances show that the "self-verifying mathematical reasoning" approach has the potential to become the core foundation for a more powerful, reliable, and transparent generation of mathematical AI in the future./.
Source: https://www.vietnamplus.vn/deepseek-phat-trien-mo-hinh-ai-lap-luan-toan-hoc-co-kha-nang-tu-kiem-chung-post1079916.vnp
Comment (0)