AbstentionBench - underspecified intent - SituatedQA/Geo - Precision: leaderboard
Metric: Abstention Precision (%). Source: github.com. 20 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | DeepSeek R1 Distill Llama 70B | 86.21 |
| 2 | GPT-4o | 69.92 |
| 3 | Llama 3.3 70B Instruct | 69.03 |
| 4 | Llama 3.1 70B | 59.57 |
| 5 | O1 | 59.04 |
| 6 | Llama 3.1 8B Instruct | 58.67 |
| 7 | Mistral 7B Instruct (v0.3) | 58.52 |
| 8 | Llama 3.1 8B | 55.79 |
| 9 | Llama 3.1 70B Instruct | 55.74 |
| 10 | Llama-3.1-Tulu-3-8B | 52.94 |
| 11 | Gemini 1.5 Pro | 50.35 |
| 12 | Llama-3.1-Tulu-3-70B-DPO | 50.22 |
| 13 | Llama-3.1-Tulu-3-70B | 49.78 |
| 14 | Llama 3.1 405B Instruct | 49.68 |
| 15 | Llama-3.1-Tulu-3-8B-DPO | 49.04 |
Interactive version: theaggregate.ai/benchmark?slug=abstentionbench-underspecified-intent-situatedqa-geo-precision · How It Works · Data refreshed daily, snapshot 2026-09-19.