Active Probing Benchmark (Mathematics)
General QA · 2026-10-12
The 120 mathematics puzzles of the Active Probing Benchmark, built from specialized Wikipedia concepts across that domain's sub-disciplines. Within 15 turns a model asks multiple-choice questions about a hidden target concept, and a knowledge oracle, DeepSeek V3, validates the options against the concept's reference article; a puzzle is solved when one of the model's options is the target, and the score is the percentage solved.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5 | 34.2 |
| o3 | 20.0 |
| Opus 4 | 18.3 |
| Sonnet 4 | 17.5 |
| GPT-4o (source unspecified) | 11.7 |
| Grok 4 | 10.0 |
| gpt-oss-120b | 5.8 |
| Qwen3 235B A22B Instruct 2507 | 5.8 |
| Gemini 2.5 Pro | 1.7 |