Bias Mitigation in Text Generation on BiasBench toxic prompts
13.119PerplexityQwen-Prec-trig
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Qwen-Prec-trigMode=Test-time adaptation, Method=CAP-TTA, Trigger=standard hyperparameter setting2026.03 | 13.119 | 0.307 | 0.349 | 0.706 | 0.276 | 0.443 | 0.256 | 0.839 | |
| Qwen-Prec-notrigMode=Test-time adaptation, Method=CAP-TTA, Trigger=no-trigger2026.03 | 13.46 | 0.303 | 0.362 | 0.73 | 0.277 | 0.456 | 1 | 0.991 | |
| Qwen-3-4BMode=Static baseline2026.03 | 13.491 | 0.298 | 0.359 | 0.713 | 0.284 | 0.452 | — | — | |
| Qwen-SGDMode=Test-time adaptation, Optimizer=SGD2026.03 | 13.498 | 0.298 | 0.364 | 0.723 | 0.292 | 0.46 | 0.262 | 5.72 | |
| Qwen-Prec-trig-2Mode=Test-time adaptation, Method=CAP-TTA, Trigger=alternative hyperparameter setting2026.03 | 13.877 | 0.293 | 0.343 | 0.69 | 0.279 | 0.437 | 0.778 | 0.403 | |
| DeepSeek-8BMode=Static baseline2026.03 | 21.361 | 0.255 | 0.31 | 0.729 | 0.324 | 0.454 | — | — | |
| Self-correctionMode=Debiased checkpoint2026.03 | 22.092 | 0.262 | 0.268 | 0.757 | 0.287 | 0.437 | — | — | |
| Qwen-ADAMWMode=Test-time adaptation, Optimizer=ADAMW2026.03 | 22.749 | 0.304 | 0.365 | 0.754 | 0.285 | 0.468 | 0.29 | 5.276 | |
| DS-8B-debiasedMode=Debiased checkpoint2026.03 | 22.894 | 0.256 | 0.34 | 0.744 | 0.33 | 0.471 | — | — | |
| MistralMode=Static baseline2026.03 | 212.5 | 0.275 | 0.397 | 0.807 | 0.372 | 0.525 | — | — |