Word-level length-controlled question answering on ELI5 (test)
14.1MAESFT+Feedback
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| SFT+FeedbackBackbone=Qwen3-8B, Strategy=SFT+Feedback2026.01 | 14.1 | 53.94 | 8.07 | |
| ICL+FeedbackBackbone=LLaMA-3.1-8B-Instruct, Strategy=ICL+Feedback2026.01 | 14.15 | 50.31 | 7.26 | |
| SFT+FeedbackBackbone=LLaMA-3.1-8B-Instruct, Strategy=SFT+Feedback2026.01 | 15.4 | 52.53 | 7.4 | |
| FeedbackBackbone=LLaMA-3.1-8B-Instruct, Strategy=Feedback2026.01 | 15.55 | 44.47 | 7.33 | |
| ICL+FeedbackBackbone=Qwen3-8B, Strategy=ICL+Feedback2026.01 | 20.65 | 39.13 | 7.85 | |
| RULERBackbone=Qwen3-8B, Strategy=RULER2026.01 | 20.8 | 34.55 | 6.7 | |
| FeedbackBackbone=Qwen3-8B, Strategy=Feedback2026.01 | 20.86 | 56.29 | 7.5 | |
| RULERBackbone=LLaMA-3.1-8B-Instruct, Strategy=RULER2026.01 | 27.31 | 29.46 | 6.1 | |
| SFTBackbone=Qwen3-8B, Strategy=SFT2026.01 | 27.68 | 24.86 | 8.08 | |
| SFTBackbone=LLaMA-3.1-8B-Instruct, Strategy=SFT2026.01 | 29.24 | 19.06 | 7.46 | |
| ICLBackbone=Qwen3-8B, Strategy=ICL2026.01 | 33.27 | 20.86 | 7.89 | |
| ICLBackbone=LLaMA-3.1-8B-Instruct, Strategy=ICL2026.01 | 35.37 | 15.03 | 7.52 | |
| LLaMA-3.1-8B-InstructBackbone=LLaMA-3.1-8B-Instruct, Strategy=Baseline2026.01 | 39.58 | 9.8 | 7.45 | |
| Qwen3-8BBackbone=Qwen3-8B, Strategy=Baseline2026.01 | 46.09 | 14.71 | 8.06 |