HELM EWoK - Physical Relations: leaderboard
HELM EWoK (Elements of World Knowledge) physical relations slice — a minimal-pair task testing whether a model prefers the plausible over the implausible world-model description (exact match).
Metric: EM. Source: crfm.stanford.edu. Saturation forecast: Estimated already saturated. 15 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet (20240620) | 83.61 | #299 |
| 2 | GPT-4 Turbo | 82.05 | #404 |
| 3 | Qwen 2 72B Instruct | 81.57 | #576 |
| 4 | GPT-4o Mini (2024-07-18) | 77.23 | #558 |
| 5 | GPT-4o (2024-05-13) | 76.51 | #325 |
| 6 | Llama 3 70B Instruct | 76.39 | #634 |
| 7 | Claude 3 Sonnet (20240229) | 76.02 | #677 |
| 8 | Mixtral 8x22B Instruct | 75.78 | #728 |
| 9 | Claude 3 Opus (20240229) | 75.18 | #412 |
| 10 | Claude 3.5 Haiku (20241022) | 71.2 | #588 |
| 11 | Mixtral 8x7B Instruct | 70.24 | #1170 |
| 12 | Claude 3 Haiku (20240307) | 70.12 | #771 |
| 13 | GPT-3.5 Turbo (0125) | 67.71 | #984 |
| 14 | Mistral 7B Instruct (v0.3) | 63.61 | #1274 |
| 15 | Llama 3 8B Instruct | 1.08 | #1129 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-ewok-physical-relations · How It Works · Data refreshed daily, snapshot 2026-10-20.