Credit score: Nikolas Kokovlis/NurPhoto through Getty
Chinese language synthetic intelligence firm DeepSeek has launched a mathematical reasoning mannequin that may establish and proper its personal errors. The mannequin beat the perfect human rating in one of many world’s most prestigious undergraduate maths competitions.
The mannequin, DeepSeekMath-V2, scored 118 out of 120 factors on questions from the 2024 William Lowell Putnam Mathematical Competitors, beating the highest human rating of 90. The mannequin additionally carried out on the degree of gold-medal winners within the Worldwide Mathematical Olympiad (IMO) 2025 and the 2024 China Mathematical Olympiad. The outcomes are described in a preprint1 posted on arXiv on 27 November.
“We’re at some extent the place AI is about pretty much as good at maths as a sensible undergraduate scholar,” says Kevin Buzzard, a mathematician at Imperial School London. “It is extremely thrilling.”
In February, AlphaGeometry 2, an AI downside solver created by Google DeepMind in London, additionally achieved a gold-level efficiency within the IMO. The feat was repeated in July by Gemini’s Deep Assume, which is owned by DeepMind.
Reasoning over solutions
Early approaches to coaching giant language fashions for mathematical reasoning centered on the accuracy of ultimate solutions, the preprint authors write. However an accurate reply doesn’t assure right reasoning. At occasions, an accurate remaining reply would possibly simply be a results of a lucky error. Furthermore, an unique give attention to the tip outcome isn’t helpful in proving mathematical legal guidelines or formulae, when the logical reasoning is extra necessary than the ultimate reply.
Tong Xie, a chemist specialising in AI-driven discoveries at UNSW Sydney in Australia, says the researchers behind DeepSeek, in addition to these growing Gemini’s Deep Assume, have been engaged on overcoming this downside by rewarding reasoning over the ultimate reply.
DeepSeekMath-V2 introduces self-verifiable mathematical reasoning for the primary time. The mannequin consists of a verifier educated to guage mathematical proofs — that are constructed on a sequence of step-by-step deductions — to establish logical flaws and assign scores based mostly on how rigorous the proof was. A meta-verification system then checks whether or not the verifier’s critiques are correct, lowering the probability of hallucinations and bettering trustworthiness. These elements work with a proof generator that constructs options and evaluates its personal work, refining arguments till no additional points could be discovered.
The design creates a suggestions loop: the verifier improves the generator, and because the generator produces more-challenging proofs, these change into new coaching information to strengthen the verifier.
The system was capable of remedy 5 out of six issues, scoring 83.3%, within the 2025 IMO. It was, nevertheless, unable to resolve the toughest issues set in 2025 and in previous IMOs.
Math-V2 depends on self-verification utilizing pure language within the mannequin itself, Xie says. This reduces human involvement and makes the mannequin cheaper and scalable.
Gemini’s Deep Assume, against this, verifies mathematical reasoning utilizing an exterior, symbolic language referred to as Lean, and its verification course of requires intensive skilled enter. The strategy is sort of freed from hallucination, however it’s computationally costly and resource-intensive, Xie says.