GPT-Erdos – GPT 5.2 attempts to solve Erdos problems, reviewed by mathematicians (github.com)

GPT-Erdos, leveraging the capabilities of GPT 5.2, presents a pioneering initiative to address and analyze the significance of Erdős problems, a renowned set of mathematical challenges. This project utilizes large language models (LLMs) for proof searching and autoformalization, aiming to uncover the strengths and weaknesses of current AI methodologies in mathematical reasoning. The initiative serves as both a benchmark for measuring mathematical capability in AI and a structured platform for examining how LLMs manage complex problem-solving, highlighting their ability to differentiate valid proof constructs and identify hidden constraints. The significance of GPT-Erdos lies in its dual role as a research tool and a performance evaluator for LLMs in mathematical theory. By submitting Erdős problems to GPT 5.2 Pro and Deep Research, the project identifies new proofs, previously overlooked solutions, and constraints not typically noted by mathematicians. Key findings include three new proofs, four exact literature solutions that were unidentified before, and significant insights into how LLMs can produce valid yet unrefined proofs. The feedback from mathematician reviewers provides additional context to these results, reinforcing GPT-Erdos' value in advancing our understanding of AI's potential in mathematics while also exposing the nuanced limitations of current AI technologies.