Hallucination Detection on WHAM
0.03FPRAudioSAE Steering
Evaluation results
| Method | Links | |
|---|---|---|
| AudioSAE SteeringAlpha (α)=3, Features=Top-100 from FSD50k2026.02 | 0.03 | |
| AudioSAE SteeringAlpha (α)=1, Features=Top-100 from FSD50k2026.02 | 0.14 | |
| S-Vector SteeringAlpha (α)=3, Calculation Dataset=Musan2026.02 | 0.32 | |
| Whisper + AudioSAESteering Method=None, Configuration=Injected SAE on last layer2026.02 | 0.51 |