AI Models Achieve 95% Accuracy on Number Theory Problems
A new paper on arXiv (2504.19451v3) demonstrates two applications of artificial intelligence to number theory. The first part evaluates the open-source LLM Qwen2.5-Math-7B-Instruct on algorithmic and computational tasks from classical textbooks and Math StackExchange. With optimal non-spoiling hints, the model achieves at least 0.95 accuracy on all thirty algorithmic problems and thirty computational questions. The second part empirically verifies a folklore conjecture in analytic number theory using ensemble methods. The research focuses on specialized domains rather than general theorem proving.
Key facts
- Paper arXiv:2504.19451v3 presents AI applications to number theory.
- Evaluates Qwen2.5-Math-7B-Instruct on algorithmic and computational tasks.
- Benchmark includes 30 algorithmic problems and 30 computational questions.
- Model achieves at least 0.95 accuracy with optimal non-spoiling hints.
- Second part verifies a folklore conjecture in analytic number theory.
- Uses ensemble methods for conjecture verification.
- Focuses on specialized domain rather than general theorem proving.
- Sources include classical textbooks and Math StackExchange.
Entities
Institutions
- arXiv