URIAL-Bench - STEM: leaderboard
The "STEM" subtask of URIAL-Bench. A task from URIAL-Bench, which evaluates base (untuned) language models aligned purely through in-context prompting (Untuned LLMs with Restyled In-context ALignment), measuring instruction-following and helpfulness without fine-tuning.
Metric: Judge Score (0-10). Source: huggingface.co. Saturation forecast: Estimated already saturated. 19 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4 | 9.7 | #438 |
| 2 | Llama 2 70B Base | 8.75 | #1025 |
| 3 | GPT-3.5 Turbo | 8.7 | #865 |
| 4 | DBRX | 8.43 | #1045 |
| 5 | Mistral-7B-v0.1 | 8.28 | #1326 |
| 6 | Mixtral 8x7B (v0.1) | 8.23 | #1083 |
| 7 | Yi 34B (Base) | 7.83 | #972 |
| 8 | Yi 6B (Base) | 7.7 | #1472 |
| 9 | gemma-7B | 7.25 | #1384 |
| 10 | Llama 2 13B Base | 7.03 | #1445 |
| 11 | Phi-2 | 7 | #1511 |
| 12 | Llama 2 7B Base | 6.8 | #1518 |
| 13 | gemma-2B | 5.73 | #1575 |
| 14 | falcon-7B | 4.43 | #1598 |
| 15 | mpt-7B | 2.25 | #1619 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=urial-bench-stem · How It Works · Data refreshed daily, snapshot 2026-10-20.