AI Systems Edge Closer to Human Mathematicians in Problem Solving

Researchers at Google DeepMind have made significant strides in artificial intelligence, bringing AI systems closer to outperforming top human mathematicians. The AI systems demonstrated a unique performance pattern, achieving perfect scores on some problems while failing to initiate solutions on others. In a recent challenge, the systems took three days to solve one complex question, contrasting with their ability to answer simpler problems in mere seconds. The two AI systems involved in the challenge, AlphaProof and AlphaGeometry 2, employed distinct methodologies. AlphaProof, which successfully tackled three problems, combines a large language model with a reinforcement learning approach, similar to techniques used in the game of Go. Thomas Hubert, the lead on AlphaProof, emphasized the goal of merging formal and informal mathematics to enhance problem-solving capabilities. AlphaProof's training involved a vast array of math problems presented in English, allowing it to generate specific proofs in formal mathematical language. While this method proved effective for complex problems, it was not always swift, as evidenced by the three-day duration required to identify the correct formal model for one of the challenge's hardest questions. In a separate context, AlphaGeometry 2 faced a problem involving a character named Turbo navigating a grid with hidden monsters. Turbo's task was to identify the location of these monsters based on specific movement rules, showcasing the diverse applications of AI in tackling mathematical and logical challenges.