AIMO Remarks - AI4Math in 2025: Closing Gaps, Exposing Fault Lines

This talk surveys recent progress within the AI4Math field, focusing on the top-performing submissions from the 2025 AI athematical Olympiad (AIMO) — one of Kaggle’s largest competitions to date — and the broader evolution of the field beyond Olympiad-style problem solving. We present two key findings: First, where consensus exists on what the right benchmark tasks are to track performance, open-source models are rapidly closing the gap with proprietary systems. Second, in areas where benchmark consensus is lacking — reflecting the sparsity of evaluation standards in AI4Math — both open-source and proprietary LLMs exhibit comparable limitations.