Personalized Response Generation on LS1 -> LS2 (test)
4.032RSAPCogAlign
Evaluation Results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| PCogAlignBackbone=Qwen2.5-VL-3B-Instruct, Strategy=SFT2025.06 | 4.032 | 4.053 | 4.211 | 4.871 | 3.963 | 4.226 | 59.3 | 26.4 | 14.3 | |
| PCogAlignBackbone=Qwen2.5-VL-3B-Instruct, Strategy=Final Selection2025.06 | 4.029 | 4.067 | 4.205 | 4.847 | 3.963 | 4.222 | 60.5 | 26.5 | 13 | |
| PCogAlignBackbone=Qwen2.5-VL-3B-Instruct, Strategy=Prompt2025.06 | 4.028 | 4.021 | 4.183 | 4.854 | 3.967 | 4.21 | 59.9 | 25.8 | 14.3 | |
| RS PromptBackbone=Qwen2.5-VL-3B-Instruct, Strategy=SFT2025.06 | 3.981 | 4.011 | 4.151 | 4.833 | 3.935 | 4.182 | 59.6 | 27.4 | 13.1 | |
| RLAIFBackbone=Qwen2.5-VL-3B-Instruct, Strategy=SFT2025.06 | 3.978 | 3.984 | 4.153 | 4.809 | 3.926 | 4.17 | 57.9 | 26.9 | 15.3 | |
| RS PromptBackbone=Qwen2.5-VL-3B-Instruct, Strategy=Prompt2025.06 | 3.953 | 3.93 | 4.073 | 4.759 | 3.885 | 4.12 | 57.8 | 29.1 | 13.1 | |
| Self-RefineBackbone=Qwen2.5-VL-3B-Instruct, Strategy=DPO2025.06 | 3.944 | 3.936 | 4.056 | 4.751 | 3.888 | 4.115 | 58.7 | 27.4 | 13.9 | |
| RLAIFBackbone=Qwen2.5-VL-3B-Instruct, Strategy=DPO2025.06 | 3.943 | 3.919 | 4.073 | 4.764 | 3.889 | 4.118 | 57.9 | 28.4 | 13.8 | |
| PCogAlignBackbone=Qwen2.5-VL-3B-Instruct, Strategy=DPO2025.06 | 3.943 | 3.926 | 4.051 | 4.769 | 3.888 | 4.115 | 57.4 | 28.8 | 13.7 | |
| RLCDBackbone=Qwen2.5-VL-3B-Instruct, Strategy=DPO2025.06 | 3.929 | 3.927 | 4.068 | 4.754 | 3.89 | 4.113 | 58.4 | 27.6 | 14 | |
| Self-RefineBackbone=Qwen2.5-VL-3B-Instruct, Strategy=SFT2025.06 | 3.924 | 3.957 | 4.056 | 4.767 | 3.899 | 4.121 | 54.8 | 28 | 17.1 | |
| RLCDBackbone=Qwen2.5-VL-3B-Instruct, Strategy=SFT2025.06 | 3.917 | 3.949 | 4.067 | 4.717 | 3.867 | 4.103 | 57.1 | 25.7 | 17.2 | |
| BaseBackbone=Qwen2.5-VL-3B-Instruct2025.06 | 3.511 | 3.556 | 3.612 | 4.436 | 3.662 | 3.755 | 0 | 100 | 0 |