Temporal Action Grounding on LAT-Bench Chinese v1
47.2mIoULAT-Audio
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| LAT-Audio2026.04 | 47.2 | 63.7 | 49 | 32.6 | |
| LAT-AudioDownsampling Factor=1x2026.04 | 45.4 | 61 | 48.8 | 20.3 | |
| LAT-AudioAblation=w/o Stage3-RL2026.04 | 45.3 | 60.8 | 45.1 | 30.2 | |
| LAT-AudioAblation=w/o Stage1-SFT+Stage3-RL2026.04 | 42.3 | 54.7 | 43.5 | 29 | |
| LAT-AudioAblation=w/o Global Timeline2026.04 | 41.6 | 45.9 | 42.8 | 28.1 | |
| Gemini-2.5-Pro2026.04 | 40.3 | 61.3 | 48.7 | 26.1 | |
| LAT-AudioDownsampling Factor=4x2026.04 | 39.1 | 53.7 | 41.1 | 27.3 | |
| LAT-AudioAblation=w/o TWA-CoT2026.04 | 38.9 | 51.3 | 40.1 | 24.9 | |
| QA-only SFTAblation=True2026.04 | 36.4 | 49.7 | 38.6 | 21.7 | |
| Gemini-2.5-ProEvaluation Strategy=Sliding Window2026.04 | 35.8 | 49.2 | 36.1 | 23.3 | |
| Gemini-3.0-Pro2026.04 | 34.6 | 50.9 | 32.8 | 22.8 | |
| LAT-AudioDownsampling Factor=8x2026.04 | 27.1 | 39.4 | 26.6 | 15.8 | |
| Qwen3-OmniEvaluation Strategy=Sliding Window2026.04 | 22.8 | 37.1 | 22.2 | 14.4 | |
| Qwen3-Omni2026.04 | 14.8 | 21.4 | 12.4 | 7 | |
| Step-Audio-R1.1Evaluation Strategy=Sliding Window2026.04 | 8.1 | 10.9 | 8.5 | 5.1 | |
| Audio-Flamingo3Evaluation Strategy=Sliding Window2026.04 | 3.7 | 6.4 | 3.9 | 2.9 |