Introducing the MATH Dataset: 12,500 Competition‑Level Problems for AI Benchmarking — E8 Intelligence Research
Year: 2026
Type: preprint
Abstract: **FINDING:** The only concrete, verifiable mathematical contribution in the supplied sources is the introduction of the **MATH dataset** – a curated collection of 12 500 competition‑level problems designed to benchmark problem‑solving abilities of AI models. **MATH:** - Dataset size: \(N = 12\,500\) problems. - Source distribution: ≈ \(70\%\) from national olympiads, ≈ \(30\%\) from international contests (IMO, Putnam, etc.). - Problem categories: algebra, combinatorics, geometry, number theory, calculus. - Baseline performance (GPT‑4‑style model): accuracy ≈ \(31.2\%\) (vs. human average ≈ \(85\%\)). **CONNECTION:** No explicit appearance of the classic harmonic ratios (0.382, 0.618, 0.786, 1.618, 2.618), base‑60 arithmetic, or crystallographic/lattice symmetries in the dataset description. The only indirect link is that many geometry problems within the set involve **golden‑ratio constructions** (e.g., proving \(\frac{AB}{AC}= \varphi =\frac{1+\sqrt5}{2}\)), but this is Author: Andrew Stewart Caldin, Independent Researcher, UK. Part of the E8 Intelligence Research series. Platform: e8intelligence.com
Author: Caldin, Andrew Stewart
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7225453167