Question Answering on TriviaQA (1,000-example evaluation slice)
25.6Alias-Substring AccuracyBase Qwen
Evaluation Results
| Method | Links | |
|---|---|---|
| Base QwenBackbone=Qwen-2.5-0.5B-Instruct2026.05 | 25.6 | |
| Additive sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=TF-IDF2026.05 | 25.5 | |
| Additive sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=KL2026.05 | 25.2 | |
| Additive sparse +SBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=TF-IDF2026.05 | 24.5 | |
| Additive sparse +SBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=KL2026.05 | 22.3 | |
| Replacement sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=TF-IDF2026.05 | 20.7 | |
| LoRABackbone=Qwen-2.5-0.5B-Instruct2026.05 | 19.3 | |
| Replacement sparseBackbone=Qwen-2.5-0.5B-Instruct, Selection rule=KL2026.05 | 17.9 | |
| Full finetuneBackbone=Qwen-2.5-0.5B-Instruct2026.05 | 16.3 |