Long-context Text Summarization on MultiNews 128K context
29.5ROUGE-LFP16
Evaluation results
| Method | Links | ||
|---|---|---|---|
| FP16Model=Llama-3.1-70B2026.03 | 29.5 | 2,400 | |
| TTKVModel=Llama-3.1-70B2026.03 | 29.3 | 580 | |
| DiffKVModel=Llama-3.1-70B2026.03 | 28.7 | 850 | |
| KVQuantModel=Llama-3.1-70B2026.03 | 28.5 | 950 | |
| ShadowKVModel=Llama-3.1-70B2026.03 | 28.5 | 820 | |
| FP16Model=Qwen2.5-32B2026.03 | 28 | 2,300 | |
| TTKVModel=Qwen2.5-32B2026.03 | 27.9 | 540 | |
| KIVIModel=Llama-3.1-70B2026.03 | 27.7 | 1,000 | |
| DiffKVModel=Qwen2.5-32B2026.03 | 27.4 | 850 | |
| KVQuantModel=Qwen2.5-32B2026.03 | 27.2 | 920 | |
| ShadowKVModel=Qwen2.5-32B2026.03 | 27.2 | 820 | |
| FP16Model=DeepSeek-R1-14B2026.03 | 27 | 2,500 | |
| TTKVModel=DeepSeek-R1-14B2026.03 | 26.9 | 450 | |
| DiffKVModel=DeepSeek-R1-14B2026.03 | 26.4 | 880 | |
| KIVIModel=Qwen2.5-32B2026.03 | 26.2 | 980 | |
| KVQuantModel=DeepSeek-R1-14B2026.03 | 26.2 | 950 | |
| ShadowKVModel=DeepSeek-R1-14B2026.03 | 26.2 | 845 | |
| KIVIModel=DeepSeek-R1-14B2026.03 | 25.3 | 1,020 |