Attribute-controlled Text Generation on HelpSteer2 Negative Representative Target Score (test)
0.987DiversityPrompting
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PromptingBackbone=Llama-3.2-3B2025.10 | 0.987 | 2.73 | 2.47 | |
| Re-ControlBackbone=Llama-3.2-3B2025.10 | 0.986 | 2.72 | 2.27 | |
| PromptingBackbone=Phi-4-mini2025.10 | 0.664 | 2.67 | 5.18 | |
| BaseBackbone=Phi-4-mini2025.10 | 0.659 | 2.76 | — | |
| BaseBackbone=Llama-3.2-3B2025.10 | 0.656 | 2.76 | — | |
| MAT-SteerBackbone=Llama-3.2-3B2025.10 | 0.539 | 2.57 | 5.84 | |
| Re-ControlBackbone=Phi-4-mini2025.10 | 0.494 | 2.56 | 5.8 | |
| ITIBackbone=Phi-4-mini2025.10 | 0.45 | 2.73 | 4.02 | |
| MAT-SteerBackbone=Phi-4-mini2025.10 | 0.308 | 2.86 | 8.73 | |
| ITIBackbone=Llama-3.2-3B2025.10 | 0.294 | 2.69 | 5.48 | |
| PRE-CONTROLBackbone=Phi-4-mini2025.10 | 0.291 | 2.46 | 9.11 | |
| PRE-CONTROLBackbone=Llama-3.2-3B2025.10 | 0.251 | 2.63 | 6.6 |