AA-Omniscience evaluates factual recall and hallucination across 6,000 questions spanning law, health, business, software engineering, humanities, and STEM — rewarding accuracy while penalizing incorrect answers and hallucinations, with no penalty for abstention.
Large-scale factuality benchmark (6K questions) across economically relevant domains. Penalizes hallucination directly rather than just measuring accuracy. Rewards models that refuse to answer when uncertain. Run independently by Artificial Analysis on dedicated hardware.
Index ranges from -100 to 100 which is unusual. Newer benchmark with limited historical data. Question generation used LLMs, so there may be subtle biases in question framing.
Higher is better. Index from -100 to 100, where 0 means equal correct and incorrect answers. Positive scores indicate more correct than incorrect responses.
Max score: 100
Index ranges from -100 to 100 which is unusual.
| # | Model | Score |
|---|---|---|
| 1 | 53.5 | |
| 2 | 48.1 | |
| 3 | 46.3 | |
| 4 | 43.7 | |
| 5 | 43.4 | |
| 6 | 43.3 | |
| 7 | 42.6 | |
| 8 | 42.4 | |
| 9 | 42.4 | |
| 10 | 42.4 | |
| 11 | 42.2 | |
| 12 | 42.0 | |
| 13 | 37.1 | |
| 14 | 35.4 | |
| 15 | 33.7 | |
| 16 | 31.9 | |
| 17 | 31.2 | |
| 18 | 30.6 | |
| 19 | 30.5 | |
| 20 | 30.5 | |
| 21 | 29.9 | |
| 22 | 29.6 | |
| 23 | 28.8 | |
| 24 | 27.2 | |
| 25 | 27.1 | |
| 26 | 26.5 | |
| 27 | 26.2 | |
| 28 | 25.8 | |
| 29 | 25.3 | |
| 30 | 25.1 | |
| 31 | 25.0 | |
| 32 | 25.0 | |
| 33 | 23.3 | |
| 34 | 23.1 | |
| 35 | 23.1 | |
| 36 | 22.7 | |
| 37 | 22.1 | |
| 38 | 22.0 | |
| 39 | 20.1 | |
| 40 | 19.7 | |
| 41 | 18.3 | |
| 42 | 18.0 | |
| 43 | 16.4 | |
| 44 | 16.4 | |
| 45 | 15.3 | |
| 46 | 13.5 | |
| 47 | 13.5 | |
| 48 | 13.5 | |
| 49 | 13.4 | |
| 50 | 13.4 |