HorizonMath Benchmark for AI Mathematical Discovery

HorizonMath is a benchmark for measuring AI progress in mathematical discovery through automated verification, as described in a 2026 arXiv paper by Erik Y. Wang and colleagues. The dataset was created by 'squashenthus' and last updated on Hugging Face in March 2026. It focuses on evaluating AI systems' ability to generate and verify mathematical statements.

Use Cases

• Benchmarking AI models on generating conjectures from mathematical premises.
• Evaluating automated verification systems on proving or disproving mathematical statements.
• Training sequence-to-sequence models for translating natural language math problems into formal logic.
• Analyzing the difficulty progression of problems across defined 'horizons' or complexity levels.

Strengths

• Based on peer-reviewed 2026 arXiv preprint (arXiv:2603.15617).
• Designed with a focus on automatic verification of results.

Limitations

• Unknown dataset size, row count, and specific problem count.
• Unknown license terms for usage and redistribution.
• Potential bias towards the specific mathematical domains covered in the paper.

Provenance

Source

Hugging Face dataset uploaded by 'squashenthus', based on the associated research paper.

Collection Method

Created for the specific research purpose outlined in the arXiv paper, methodology unknown.

Time Range

null

Freshness

Last updated on the platform in March 2026.

Geography

null