Audio Reasoning on MMAR N=1,000
53.6AccuracyCoFi-Agent
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| CoFi-Agentvariant=Adaptive + ASR2026.01 | 53.6 | 9.617 | |
| CoFi-Agentvariant=Always-On + ASR2026.01 | 51.7 | 11.058 | |
| CoFi-Agentvariant=Adaptive Re-listen2026.01 | 43.1 | 6.446 | |
| Hybridvariant=Describe→Reason2026.01 | 39.3 | 2.866 | |
| Qwen2-Audiovariant=Baseline2026.01 | 27.2 | 0.155 |