Long-context Language Modeling on LongBench-En Filtered Tasks
42.6Single Document ScoreCPC
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| CPCProxy Model=Mistral-7B2025.05 | 42.6 | 48.6 | 23.7 | 38.3 | 1,844 | 5 | |
| SentinelProxy Model=Qwen2.5-1.5B2025.05 | 40.6 | 48.1 | 26 | 38.2 | 1,883 | 5 | |
| SentinelProxy Model=Qwen2.5-0.5B2025.05 | 40.1 | 47.4 | 25.8 | 37.8 | 1,885 | 5 | |
| Original Prompt2025.05 | 39.7 | 38.7 | 26.5 | 35 | 10,295 | — | |
| LongLLMLinguaProxy Model=LLaMA2-7B2025.05 | 39 | 42.2 | 27.4 | 36.2 | 1,809 | 6 | |
| SentinelProxy Model=Mistral-7B2025.05 | 38.3 | 47 | 24 | 36.4 | 1,873 | 5 | |
| LLMLingua-2Proxy Model=XLM-R-Large-0.6B2025.05 | 29.8 | 33.1 | 25.3 | 29.4 | 1,954 | 5 | |
| LLMLinguaProxy Model=LLaMA2-7B2025.05 | 22.4 | 32.1 | 24.5 | 26.3 | 1,950 | 5 | |
| Selective-ContextProxy Model=LLaMA2-7B2025.05 | 16.2 | 34.8 | 24.4 | 25.1 | 1,925 | 5 |