Question Answering on PubMedQA (Acc, TNFT, Tokens, Delay)
73.6AccuracyVanilla
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| VanillaBackbone=Qwen3-4B2026.05 | 73.6 | 404.67 | 1,420.14 | 21.73 | |
| OursBackbone=Qwen3-4B2026.05 | 73.5 | 0 | 648.96 | 13.8 | |
| BaseBackbone=Qwen3-4B2026.05 | 72.9 | 389.6 | 1,323.75 | 20.98 | |
| OursBackbone=Qwen3-1.7B2026.05 | 70.5 | 0 | 658.01 | 10.96 | |
| VanillaBackbone=Qwen3-1.7B2026.05 | 70.4 | 404.67 | 1,422.91 | 15.6 | |
| BaseBackbone=Qwen3-1.7B2026.05 | 69.8 | 389.6 | 1,263.91 | 14.2 |