Controlled Text Generation on Syntactic Control (Q ∝ pq) (test)
0.0001Log Probability Q(y)GPT2-large + SIS
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| GPT2-large + SISBackbone=GPT2-large, Algorithm=SIS, Few-shot setting=None2025.06 | 0.0001 | -50.57 | 91.22 | 53 | 70 | 66 | |
| GPT2-large + SMCBackbone=GPT2-large, Algorithm=SMC, Few-shot setting=None2025.06 | -0.0009 | -47.98 | 93.69 | 56 | 73 | 67 | |
| Llama3-8B (5-shot) + SMCBackbone=Llama3-8B, Algorithm=SMC, Few-shot setting=5-shot2025.06 | -0.001 | -29.96 | 93.05 | 88 | 87 | 75 | |
| Llama3-8B (0-shot) + SISBackbone=Llama3-8B, Algorithm=SIS, Few-shot setting=0-shot2025.06 | -0.002 | -32.44 | 91.58 | 71 | 75 | 67 | |
| Llama3-8B (5-shot) + SISBackbone=Llama3-8B, Algorithm=SIS, Few-shot setting=5-shot2025.06 | -0.002 | -31.87 | 92.24 | 74 | 79 | 69 | |
| Llama3-8B + SISBackbone=Llama3-8B, Algorithm=SIS, Few-shot setting=None2025.06 | -0.28 | -75.58 | 86.5 | 75 | 81 | 71 | |
| Llama3-8B (0-shot) + SMCBackbone=Llama3-8B, Algorithm=SMC, Few-shot setting=0-shot2025.06 | -0.28 | -30.36 | 92.78 | 87 | 87 | 75 | |
| Llama3-8B + SMCBackbone=Llama3-8B, Algorithm=SMC, Few-shot setting=None2025.06 | -6.68 | -70.8 | 90.18 | 88 | 86 | 73 | |
| Llama3-8B (5-shot)Backbone=Llama3-8B, Algorithm=None, Few-shot setting=5-shot2025.06 | -8.02 | -37.17 | 86.77 | 75 | 79 | 70 | |
| GPT2-largeBackbone=GPT2-large, Algorithm=None, Few-shot setting=None2025.06 | -12.71 | -54.91 | 82.89 | 54 | 72 | 67 | |
| Llama3-8BBackbone=Llama3-8B, Algorithm=None, Few-shot setting=None2025.06 | -14.93 | -89.11 | 76.49 | 75 | 81 | 71 | |
| Llama3-8B (0-shot)Backbone=Llama3-8B, Algorithm=None, Few-shot setting=0-shot2025.06 | -15.83 | -39.16 | 86.38 | 72 | 76 | 68 |