Video Temporal Grounding on QVHighlights zero-shot transfer
29.7mAP (zero-shot)Qwen3-VL-8B + F2G-FT
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Qwen3-VL-8B + F2G-FTBackbone=Qwen3-VL-8B-Instruct, Evaluation Protocol=Foresee-to-Ground fine-tuning2026.05 | 29.7 | 45.6 | |
| TRACEBackbone=CLIP ViT-L + Mistral-7B, Evaluation Protocol=Zero-shot transfer2026.05 | 26.8 | 42.7 | |
| NumProBackbone=LongVA-7B-DPO, Evaluation Protocol=Zero-shot transfer2026.05 | 25 | 37.2 | |
| Qwen3-VL-8B + FTBackbone=Qwen3-VL-8B-Instruct, Evaluation Protocol=Fine-tuning2026.05 | 24.6 | 36.8 | |
| Qwen3-VL-8BBackbone=Qwen3-VL-8B-Instruct, Evaluation Protocol=Zero-shot transfer2026.05 | 21.3 | 32.6 | |
| VTG-LLMBackbone=ViT-G/14 + BLIP + LLaMA-2-7B, Evaluation Protocol=Zero-shot transfer2026.05 | 16.5 | 33.5 | |
| TimeChatBackbone=ViT-G/14 + LLaMA-2-7B, Evaluation Protocol=Zero-shot transfer2026.05 | 14.5 | 23.9 | |
| MomentorBackbone=CLIP ViT-L4 + LLaMA-7B, Evaluation Protocol=Zero-shot transfer2026.05 | 7.6 | — |