Question Answering on LongBench key_facts v2
100AccuracyGPT-4.1
Evaluation results
| Method | Links | ||
|---|---|---|---|
| GPT-4.1Compression Method=Original2026.05 | 100 | — | |
| GPT-4.1 (TE)Compression Method=Telegraph English, Compression Ratio=~50%2026.05 | 99.1 | -0.9 | |
| GPT-4o-miniCompression Method=Original2026.05 | 99.1 | — | |
| GPT-4.1 (LLML2-50)Compression Method=LLMLingua-2, Retention Rate=50%2026.05 | 99 | -1 | |
| GPT-4.1-nanoCompression Method=Original2026.05 | 98 | — | |
| GPT-4o-mini (TE)Compression Method=Telegraph English, Compression Ratio=~50%2026.05 | 95.7 | -3.4 | |
| GPT-4.1-nano (TE)Compression Method=Telegraph English, Compression Ratio=~50%2026.05 | 95 | -3 | |
| GPT-4.1-nano (LLML2-50)Compression Method=LLMLingua-2, Retention Rate=50%2026.05 | 94.9 | -3.1 | |
| GPT-4o-mini (LLML2-50)Compression Method=LLMLingua-2, Retention Rate=50%2026.05 | 94.6 | -4.5 |