NLG (Table-to-Text, Summarization) on Generation OOD
34.1Score (Full Output)Qwen3-14B Base
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-14B BaseModel size=14B2026.05 | 34.1 | 31.9 | 20.3 | |
| Qwen3-8B BaseModel size=8B2026.05 | 28.4 | 24.7 | 19.4 | |
| Qwen3-8B GRPOModel size=8B, Optimization method=GRPO2026.05 | 28.4 | 24.9 | 19.2 | |
| Qwen3-8B CCOPDModel size=8B, Optimization method=CCOPD2026.05 | 27.9 | 23.3 | 22.9 | |
| Qwen3-8B 14B-teacher CCOPDModel size=8B, Optimization method=CCOPD, Teacher source=14B-teacher2026.05 | 27.8 | 23.9 | 20.6 | |
| Qwen3-8B OPSDModel size=8B, Optimization method=OPSD2026.05 | 27.7 | 23.8 | 17.7 | |
| Qwen3-8B SFTModel size=8B, Optimization method=SFT2026.05 | 27.4 | 22.4 | 19 | |
| Qwen3-8B Forward-CCOPDModel size=8B, Optimization method=Forward-CCOPD2026.05 | 27 | 23.3 | 20.4 | |
| Qwen3-8B 4B-teacher CCOPDModel size=8B, Optimization method=CCOPD, Teacher source=4B-teacher2026.05 | 26.4 | 23.2 | 21.1 | |
| Qwen3-4B BaseModel size=4B2026.05 | 22.4 | 22.6 | 11.5 | |
| Qwen3-4B CCOPDModel size=4B, Optimization method=CCOPD2026.05 | 21 | 18.4 | 17.2 | |
| Llama3.1-8B CCOPDModel size=8B, Optimization method=CCOPD2026.05 | 14.8 | 13.8 | 13.4 | |
| Llama3.1-8B BaseModel size=8B2026.05 | 13 | 16.3 | 13.1 |