End of saturation: AMO-Bench

The saturation problem

Why a new benchmark is needed: existing math benchmarks such as AIME24 and AIME25 face severe performance saturation.

Recent results show that leading models can exceed 90% accuracy on these tests. Such scores reduce the ability of these benchmarks to distinguish further progress.

AMO-Bench: advanced math reasoning

To address this, researchers introduce AMO-Bench, an advanced math reasoning benchmark with 50 carefully crafted problems.

Key properties ensure strict evaluation:

• Very high difficulty: experts cross validate every problem to reach or exceed International Mathematical Olympiad level.
• Fully original: all 50 problems are newly created to prevent inflated scores from memorizing existing contest sets.
• Automatic grading: each problem requires only a final answer rather than a full proof, enabling scalable automatic scoring with consistent evaluation.

Key finding: models still struggle

Under the stricter AMO-Bench setting, the apparent advantage of current large language models fades quickly.

Results show that the best performing model GPT-5 Thinking High reaches only 52.4% accuracy.

Most models score below 40% on this benchmark.

These results indicate substantial room for improvement in complex mathematical reasoning.