MathNet opens Olympiad math benchmark

MathNet is less a dataset drop than a stress test for the next wave of reasoning models: if models can memorize popular math sets, this gives evaluators a broader, messier global corpus to probe real generalization.

• –The retrieval angle matters because math RAG fails when embeddings match surface wording instead of proof structure or mathematical equivalence
• –Strong models still leave headroom on the benchmark, which makes this useful for measuring progress beyond saturated grade-school math tests
• –The multilingual and diagram-heavy coverage should expose weaknesses in multimodal reasoning, OCR pipelines, and non-English mathematical notation
• –Public access gives smaller labs a serious eval corpus without needing to scrape scattered Olympiad archives themselves