LongBench v2 measures long-context understanding across tasks from single-document QA to repository and structured-data analysis. It targets context lengths from 8K up to roughly 2M tokens.
Coverage spans multiple long-context regimes, including multi-doc reasoning and code-repo comprehension; objective MCQ scoring improves repeatability across runs.
MCQ framing can overestimate capability versus open-ended generation tasks; context packing strategy and CoT settings can significantly change results.
Higher is better. Accuracy percentage on MCQ tasks.
Max score: 100
MCQ framing can overestimate capability versus open-ended generation tasks.
| # | Model | Score |
|---|---|---|
| 1 | 62.4 | |
| 2 | 54.7 | |
| 3 | 54.6 | |
| 4 | 51.4 | |
| 5 | 50.4 | |
| 6 | 50.4 | |
| 7 | 50.1 | |
| 8 | 50.1 | |
| 9 | 50.1 | |
| 10 | 49.1 | |
| 11 | 48.9 | |
| 12 | 48.7 | |
| 13 | 48.1 | |
| 14 | 47.4 | |
| 15 | 47.4 | |
| 16 | 47.4 | |
| 17 | 46.9 | |
| 18 | 46.7 | |
| 19 | 46.7 | |
| 20 | 46.5 | |
| 21 | 46.5 | |
| 22 | 46.3 | |
| 23 | 46.3 | |
| 24 | 46.2 | |
| 25 | 46.2 | |
| 26 | 46.1 | |
| 27 | 46.0 | |
| 28 | 45.8 | |
| 29 | 45.7 | |
| 30 | 45.6 | |
| 31 | 45.5 | |
| 32 | 45.5 | |
| 33 | 45.4 | |
| 34 | 45.4 | |
| 35 | 45.4 | |
| 36 | 45.3 | |
| 37 | 45.3 | |
| 38 | 45.3 | |
| 39 | 45.2 | |
| 40 | 45.1 | |
| 41 | 45.1 | |
| 42 | 45.0 | |
| 43 | 45.0 | |
| 44 | 44.9 | |
| 45 | 44.9 | |
| 46 | 44.8 | |
| 47 | 44.7 | |
| 48 | 44.7 | |
| 49 | 44.6 | |
| 50 | 44.5 |