Semantic Trap on OpenAgent Tier-1
28.6Accuracy DeltaSFT-200 + PAFT
Evaluation Results
| Method | Links | |
|---|---|---|
| SFT-200 + PAFTtraining_steps=200, PAFT=true2026.07 | 28.6 | |
| SFT-400 + PAFTtraining_steps=400, PAFT=true2026.07 | 5.6 | |
| SFT-600 + PAFTtraining_steps=600, PAFT=true2026.07 | -2.5 | |
| SFT-800 + PAFTtraining_steps=800, PAFT=true2026.07 | -4.1 | |
| PAFT (alpha=0.3 default)training_steps=800, perturbation_ratio=0.32026.07 | -4.1 | |
| PAFT (alpha=0.4)training_steps=800, perturbation_ratio=0.42026.07 | -5.3 | |
| PAFT (alpha=0.2)training_steps=800, perturbation_ratio=0.22026.07 | -6.9 | |
| PAFT (SRP only)training_steps=800, perturbation_ratio=0.3, component=SRP2026.07 | -14.9 | |
| Base2026.07 | -29.8 | |
| PAFT (EFP only)training_steps=800, perturbation_ratio=0.3, component=EFP2026.07 | -30.1 | |
| PAFT (SBP only)training_steps=800, perturbation_ratio=0.3, component=SBP2026.07 | -38.9 | |
| SFT-800training_steps=8002026.07 | -50.4 | |
| SFT-600training_steps=6002026.07 | -51.3 | |
| SFT-400training_steps=4002026.07 | -53.9 | |
| SFT-200training_steps=2002026.07 | -67.7 |