Multi-turn conversation on MT-Bench-101
8.5Grounding ScoreTCA-Attention
Evaluation Results
| Method | Links | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| TCA-AttentionBackbone=Qwen2.5-7B-Instruct2025.12 | 8.5 | 7.71 | 9.56 | 9.74 | 7.48 | 9.91 | 9.39 | 9.49 | 9.26 | 8.38 | 9.2 | 8.62 | 9.43 | 8.97 | |
| Qwen2.5-7B-InstructBackbone=Qwen2.5-7B-Instruct, Attention=Full-Attention2025.12 | 8.2 | 7.71 | 9.49 | 9.56 | 7.57 | 9.89 | 8.9 | 9.5 | 9.29 | 8.26 | 9.19 | 8.64 | 9.42 | 8.9 | |
| FlexPrefillBackbone=Qwen2.5-7B-Instruct2025.12 | 7.81 | 7.67 | 9.55 | 9.57 | 7.23 | 9.91 | 9.4 | 9.57 | 9.22 | 8.33 | 9.27 | 8.63 | 9.53 | 8.9 |