Reasoning Temporal Localization on ReXTime
52.88R@0.3EventCoT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| EventCoTBackbone=Qwen2.5-7B, # Visual Tokens=50–100 (89 on average), Evaluation Protocol=Zero-shot2026.07 | 52.88 | 38.65 | 40.88 | 68.7 | 28.4 | |
| EventCoTBackbone=Vicuna-7B, # Visual Tokens=51–100 (81 on average), Evaluation Protocol=Zero-shot2026.07 | 46.8 | 33.33 | 36.5 | 48.21 | 18.1 | |
| GraphThinkerBackbone=Qwen2.5-VL-7B, Evaluation Protocol=Zero-shot2026.07 | 33.92 | 20.25 | 25.34 | 66.82 | 15.21 | |
| TOGABackbone=Mistral-7B, # Visual Tokens=2,028, Evaluation Protocol=Zero-shot2026.07 | 29.91 | 19.79 | 25.53 | — | — | |
| LITA-13BBackbone=Vicuna-13B, # Visual Tokens=356, Evaluation Protocol=Zero-shot2026.07 | 29.49 | 16.29 | 21.49 | 34.44 | — | |
| VTimeLLMBackbone=Vicuna-7B, # Visual Tokens=100, Evaluation Protocol=Zero-shot2026.07 | 28.84 | 17.41 | 20.14 | 36.16 | — | |
| Qwen2.5-VLBackbone=Qwen2.5-VL-7B, # Visual Tokens=≤24,576, Evaluation Protocol=Zero-shot2026.07 | 16.05 | 9.24 | 13.6 | 56.6 | 6.35 | |
| TimeChatBackbone=LLaMA2-7B, # Visual Tokens=3,072 (96), Evaluation Protocol=Zero-shot2026.07 | 14.42 | 7.61 | 11.65 | 40.04 | — |