AI Models Achieve Gold Medal in International Mathematical Olympiad, Raising Questions on Accuracy and Grading

The International Mathematical Olympiad (IMO) has been challenged by artificial intelligence (AI) models from OpenAI and Google-DeepMind. These models, using general-purpose reasoning and agents to gather information, were able to achieve Gold medal-level scores, demonstrating significant progress in mathematical problem-solving capabilities. However, some issues have been raised regarding the grading and accuracy of the AI solutions.

The International Mathematical Olympiad (IMO), one of the world's most prestigious math competitions for high school students, has been challenged by artificial intelligence (AI) models from OpenAI and Google DeepMind. These models, utilizing general-purpose reasoning and agents to gather information, achieved Gold medal-level scores, demonstrating significant progress in mathematical problem-solving capabilities. However, the grading and accuracy of the AI solutions have raised concerns.

The 2025 IMO, held in Australia's Sunshine Coast, saw AI models from OpenAI and Google DeepMind participating in a computerized approximation of the exam. OpenAI and Google DeepMind announced that their models earned unofficial gold medals for solving five out of six problems. This achievement was celebrated as a "moon landing moment" by industry researchers [1].

However, the hype around these results is not without controversy. The models' performance was not validated by the IMO, and the methodologies used, including the amount of compute and human involvement, remain unclear. Moreover, the IMO questions are complex and often require deep mathematical understanding that goes beyond what AI models have demonstrated so far [1].

Some experts caution against overstating AI's capabilities. Terence Tao, a prominent mathematician, noted that the testing methodology significantly influences what AI can achieve. Gregor Dolinar, the IMO president, echoed this sentiment, stating that the organization cannot validate the methods used by the AI models [1].

The AI models' performance also raises questions about the future of professional mathematicians. While AI has shown promise in solving complex problems, it is not yet capable of the deep, multi-year research required for frontier mathematical research. Mathematicians like Kevin Buzzard argue that AI-generated solutions, while impressive, do not replace the expertise and insight of human mathematicians [1].

OpenAI's recent launch of GPT-5, an advanced AI model with enhanced reasoning, coding, and contextual awareness, further highlights the rapid advancements in AI technology. GPT-5 has shown superior coding accuracy and task execution, setting new benchmarks in AI capabilities [2].

Despite these advancements, the IMO results underscore the need for caution in evaluating AI's impact on mathematics. While AI models can solve complex problems quickly, they often rely on "best-of-n" strategies and may not always provide accurate or rigorous solutions. Formal proof assistants, which can verify the logic of mathematical arguments, offer a more reliable approach to AI-generated proofs [1].

In conclusion, the IMO results highlight both the potential and the limitations of AI in mathematical problem-solving. While AI models have made significant strides, they are not yet capable of replacing human mathematicians. As AI continues to evolve, it is essential to maintain a balanced perspective and focus on the unique contributions that both AI and human expertise can bring to the field of mathematics.

References:
[1] https://www.scientificamerican.com/article/mathematicians-question-ai-performance-at-international-math-olympiad/
[2] https://www.ainvest.com/news/openai-unveils-gpt-5-enhanced-reasoning-usability-2508/