AI models fail to tackle complex math proofs effectively

In recent events dominating the field of artificial intelligence, researchers have assessed the performance of leading AI reasoning models on high-level mathematical challenges. This study was conducted shortly after the release of the problems from the 2025 US Math Olympiad in the United States. The results highlighted a significant performance decline, with most AI models scoring below 5 percent on average when asked to generate complete mathematical proofs. Notably, Google's Gemini 2.5 Pro achieved a higher average score of 10.1 out of 42 points, approximately 24 percent, leading to concerns about the models' capabilities. Moreover, OpenAI's o3-mini had the lowest average score at just 0.9 points, which translates to roughly 2.1 percent. The research team discovered recurring patterns of failure in the models’ responses. For instance, in one particular problem, the model correctly identified necessary conditions but ultimately arrived at an incorrect final answer. This inconsistency reveals a deeper issue: while AI models utilize chain-of-thought reasoning, they often struggle when faced with original proof challenges requiring detailed mathematical insight. The overall performance of these models dropped considerably when attempting problems that deviated from their training data. Chain-of-thought and simulated reasoning strategies improve results in standard problem-solving scenarios; however, these models still lack the depth of reasoning required for intricate mathematical proofs. As the study suggests, mere scaling of current architectures and enhancing training methods alone may not resolve these limitations in genuine reasoning abilities. In fact, earlier evaluations conducted by Hamed Mahdavi and colleagues at Pennsylvania State University corroborated these findings, indicating a persistent gap in AI models' understanding of high-level mathematical concepts. Looking ahead, there is cautious optimism that as these AI systems evolve, they may be able to bridge the reasoning gap through better connections within latent space. Accomplishing this would represent a significant breakthrough in the development of AI reasoning models, ultimately enhancing their ability to tackle complex mathematical challenges in the future.