IMO-ProofBench Advanced: USAMO 2025
Math · 2025-11-03
The USAMO 2025 breakdown of Advanced IMO-ProofBench, the hard proof-writing subset of Google DeepMind's IMO-Bench: its six problems are the unmodified USAMO 2025 problems, the same set as MathArena's USAMO 2025 track. IMO-Bench's maintainers grade each proof from 0 to 7 under their own guidelines and report awarded points as a percentage of the 42 available; the paper's results were graded by human experts, while later leaderboard rows do not state the grader. It is a separate protocol from MathArena's and is not pooled with it.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.5 Pro | 100.0 |
| Gemini 3 Deep Think | 83.3 |
| GPT-5.5 | 83.3 |
| Grok 4 Heavy | 76.2 |
| o3 | 52.4 |
| Gemini 2.5 Deep Think | 52.4 |
| GPT-5.2 | 50.0 |
| Opus 4.5 | 42.9 |
| Gemini 3.1 Pro Preview | 40.5 |
| Gemini 3 Pro Preview | 40.5 |
| GPT-5 Pro | 35.7 |
| Gemini 2.5 Pro | 33.3 |
| o4-mini | 23.8 |
| Grok 4 | 23.8 |
| Kimi K2 Instruct | 21.4 |
| GPT-5 | 19.0 |
| Grok 4.1 Fast | 16.7 |
| GPT-5.1 | 16.7 |
| Opus 4 | 11.9 |
| Qwen3 235B A22B Instruct 2507 | 4.8 |
| Sonnet 4 | 2.4 |
| DeepSeek-V3-0324 | 0.0 |
| DeepSeek-R1-0528 | 0.0 |