Active Probing Benchmark (APB)
General QA · 2026-10-12
960 puzzles built from specialized, doctorate-level Wikipedia concepts, 120 in each of biology, human disease, chemistry, physics, mathematics, computer science, economics and philosophy. 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 | 33.4 |
| Opus 4 | 27.6 |
| Sonnet 4 | 26.4 |
| o3 | 25.3 |
| GPT-4o (source unspecified) | 19.6 |
| Grok 4 | 17.6 |
| gpt-oss-120b | 13.4 |
| Qwen3 235B A22B Instruct 2507 | 12.6 |
| Gemini 2.5 Pro | 12.5 |