Tests commonsense physical reasoning by asking models to complete everyday activity descriptions. Created by adversarially filtering to find examples that fool earlier models.
This definition has not completed editorial and provenance review. Verify its metric, version, source, and evaluation configuration upstream before using it for a model-selection decision.
Adversarial construction provides meaningful difficulty. Tests intuitive physical reasoning.
Largely saturated for frontier models. Non-standard score representation (range 0–10). In training data for most models.
Higher is better. Score range 0–10 (Epoch-specific encoding).
Largely saturated for frontier models.
| # | Model | Score |
|---|---|---|
| 1 | 96.7 | |
| 2 | 76.8 | |
| 3 | 73.9 | |
| 4 | 71.9 | |
| 5 | 67.3 | |
| 6 | 65.1 | |
| 7 | 63.5 | |
| 8 | 63.2 | |
| 9 | 59.6 | |
| 10 | 58.3 | |
| 11 | 54.9 | |
| 12 | 51.8 | |
| 13 | 50.9 | |
| 14 | 45.2 | |
| 15 | 43.1 | |
| 16 | 42.9 | |
| 17 | 42.0 | |
| 18 | 41.0 | |
| 19 | 40.2 | |
| 20 | 38.3 | |
| 21 | 36.9 | |
| 22 | 36.0 | |
| 23 | 35.5 | |
| 24 | 35.2 | |
| 25 | Gemini 2.5 Flash-Lite Preview (Sep '25) (Non-reasoning) source benchpress Evidence observed Apr 13, 2026 | 33.7 |
| 26 | 30.7 | |
| 27 | 30.1 | |
| 28 | 29.8 | |
| 29 | 28.6 | |
| 30 | 26.8 | |
| 31 | 26.5 | |
| 32 | 26.5 | |
| 33 | 25.8 | |
| 34 | 24.6 | |
| 35 | 23.9 | |
| 36 | 23.9 | |
| 37 | 21.6 | |
| 38 | 21.5 | |
| 39 | 20.0 | |
| 40 | 20.0 | |
| 41 | 20.0 | |
| 42 | 19.9 | |
| 43 | 19.6 | |
| 44 | 19.4 | |
| 45 | 19.3 | |
| 46 | 19.0 | |
| 47 | 18.0 | |
| 48 | 17.5 | |
| 49 | 16.7 | |
| 50 | 15.5 |