LLM Agent Team Solves Complex Mathematical Research Problems.

Mathematical research presents significant hurdles for AI, characterized by its complex, non-linear reasoning paths and strict logical demands. To address this, researchers have developed the MechMath Agent Team (MMAT), an AI system powered by large language models, intended to assist throughout the entire mathematical research process. MMAT employs a "Harness Architecture" that separates system responsibilities into three planes: Control, Execution, and Augmentation. This design allows for both rigorous logical oversight and the flexibility needed for open-ended research. Within this framework, three specialized agents operate in a closed loop: a Knowledge Base Manager, a Natural Language Prover, and a Formal Language Prover, working together to generate formally certified mathematical proofs. Over a two-month evaluation period, MMAT successfully solved 11 open problems across various fields of mathematics, including Number Theory and Algebraic Complexity Theory. This demonstrates its potential as a powerful co-pilot for human mathematicians, significantly advancing AI's capability in complex reasoning tasks.