Audio Multiple-Choice Question Answering on MMAU mini 1,000 examples (test)
69.4AccuracyHead-guided layer steering
Evaluation Results
| Method | Links | |
|---|---|---|
| Head-guided layer steeringModel=R1-AQA2026.03 | 69.4 | |
| Head-level steering (specialist)Model=R1-AQA2026.03 | 67.8 | |
| Best single layer steeringModel=R1-AQA2026.03 | 65.21 | |
| Baseline (no intervention)Model=R1-AQA2026.03 | 64.5 | |
| Head-level steering (random heads)Model=R1-AQA2026.03 | 64.02 | |
| Head-guided layer steeringModel=Qwen2-Audio2026.03 | 57.25 | |
| Head-level steering (specialist)Model=Qwen2-Audio2026.03 | 54.3 | |
| Best single layer steeringModel=Qwen2-Audio2026.03 | 53.82 | |
| Head-level steering (random heads)Model=Qwen2-Audio2026.03 | 49.42 | |
| Baseline (no intervention)Model=Qwen2-Audio2026.03 | 49.2 |