Temporal Action Grounding on LAT-Bench English v1
50mIoULAT-Audio
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| LAT-Audio2026.04 | 50 | 68.1 | 54.1 | 34.5 | |
| LAT-AudioDownsampling Factor=1x2026.04 | 48.7 | 64.8 | 52 | 30.1 | |
| LAT-AudioAblation=w/o Stage3-RL2026.04 | 47.3 | 68 | 50.3 | 31.1 | |
| Gemini-2.5-Pro2026.04 | 45.3 | 65.2 | 53.9 | 27.7 | |
| LAT-AudioAblation=w/o Global Timeline2026.04 | 45.3 | 52.8 | 48.4 | 29.5 | |
| LAT-AudioAblation=w/o Stage1-SFT+Stage3-RL2026.04 | 45.2 | 60.3 | 52.9 | 27.9 | |
| LAT-AudioDownsampling Factor=4x2026.04 | 41.5 | 61.9 | 41.6 | 24.2 | |
| Gemini-3.0-Pro2026.04 | 41 | 51.4 | 44.8 | 22.9 | |
| Gemini-2.5-ProEvaluation Strategy=Sliding Window2026.04 | 40.6 | 54 | 42.7 | 29.7 | |
| LAT-AudioAblation=w/o TWA-CoT2026.04 | 40.3 | 55 | 43.8 | 26.7 | |
| QA-only SFTAblation=True2026.04 | 39.2 | 54.2 | 41.7 | 24.5 | |
| LAT-AudioDownsampling Factor=8x2026.04 | 27.8 | 45.2 | 29.8 | 15.7 | |
| Qwen3-OmniEvaluation Strategy=Sliding Window2026.04 | 26.2 | 41.9 | 25.8 | 15.5 | |
| Qwen3-Omni2026.04 | 15.8 | 26.4 | 16 | 7 | |
| Step-Audio-R1.1Evaluation Strategy=Sliding Window2026.04 | 9 | 11.2 | 9.3 | 6.1 | |
| Audio-Flamingo3Evaluation Strategy=Sliding Window2026.04 | 4.1 | 7 | 4.3 | 2.9 | |
| Time-AudioEvaluation Strategy=Sliding Window2026.04 | 2.5 | 2.9 | 2.3 | 1.1 |