AdvancedMathBench - VerifierBench: leaderboard
VerifierBench of AdvancedMathBench (Shanghai AI Laboratory, 2026) tests models as proof verifiers on 888 model-generated proof trajectories with expert labels. This board is the meta-verification balanced F1 in percent, where a verdict counts only when its rationale and error localization agree with the expert annotation; higher is better.
Metric: Meta-verification balanced F1 (%): a validity verdict counts only when its rationale and error localization agree with the expert annotation, 888 model-generated proof trajectories with expert labels; balanced F1 is the harmonic mean of the true positive and true negative rates; higher is better. Source: arxiv.org. Saturation forecast: Around 2029. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro (Max) | 65.1 | #188 (DeepSeek V4 Pro) |
| 2 | GPT-5.5 (xHigh) | 64.9 | #49 (GPT-5.5) |
| 3 | GPT-5.5 (High) | 63.6 | #49 (GPT-5.5) |
| 4 | GLM-5.2 (Max) | 63.3 | #138 (GLM-5.2) |
| 5 | Qwen 3.5 397B A17B | 58.5 | #246 |
| 6 | GPT-5.2 (xHigh) | 57.7 | #176 (GPT-5.2) |
| 7 | Gemini 3.1 Pro (Preview) (High) | 55.2 | #94 (Gemini 3.1 Pro (Preview)) |
| 8 | Claude Opus 4.8 (Max) | 51 | #56 (Claude Opus 4.8) |
| 9 | GPT-OSS-120B (High) | 47.9 | #700 (GPT-OSS-120B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=advancedmathbench-verifierbench · How It Works · Data refreshed daily, snapshot 2026-11-13.