LLM solves previously unsolved Erdos Problem by being prompted to think unconventionally—raising questions about whether reframing problems can unlock AI creativity

An unnamed LLM, when prompted to adopt a novel and creative approach, successfully solved an Erdos Problem (a famous decades-old mathematics puzzle). The breakthrough came from encouraging the model to explore non-trivial and unconventional solutions rather than following standard reasoning paths.

The key difference: instead of asking the AI to solve a problem in its default way, researchers framed the challenge so the model would actively search for creative angles. This suggests that *how* you phrase a question to an AI can be as important as *what* question you ask.

For researchers and practitioners using LLMs: this hints that current AI models may already contain problem-solving capability we're not accessing—and that prompting strategy (the exact wording and framing you give to an AI) could unlock solutions to hard problems in math, science, and engineering that previously seemed out of reach.

The discoverer proposes a next step: a "folder language" system where complex problems (like housing affordability) are broken into multiple agent perspectives (renters, buyers, policy makers), with the AI's instructions dynamically adjusted based on the problem structure being explored—but this remains a conceptual idea, not yet deployed.