VideoRFT-7B: benchmark results
VideoRFT research-team checkpoint, reinforced fine-tuned from Qwen2.5-VL-7B-Instruct after its separate SFT stage. The official model card announces the 7B weight release on May 25, 2025. This is the final RL model, not VideoRFT-SFT. Provider: Other. Released 2025-05-25. Access: Open.
Aggregate Rating 1526 ± 45, rank #1997 of 10916 rated models, from 17 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EgoPro-Bench - Action | 34.19 | F1 (%; alert timing for scene events and human actions, 200 | 50 |
| VideoZeroBench - Spatial Grounding | 4.2 | Mean visual IoU (%) of the bounding boxes the model predicts | 50 |
| VideoZeroBench - Temporal Grounding | 3.6 | Mean multi-segment temporal IoU (%) between the evidence int | 43.8 |
| EgoPro-Bench - Object | 44.15 | F1 (%; alert timing for the appearance of target objects, 20 | 37.5 |
| VideoLVR Paper - LongVideo-Reason | 64 | Accuracy (%) | 36.4 |
| VideoLVR Paper - VideoMathQA | 23.1 | Accuracy (%) | 31.8 |
| VideoLVR Paper - MMVU | 49.1 | Accuracy (%) | 27.3 |
| VideoZeroBench - Level-4 (Answer with Temporal Grounding) | 0 | Level-4 accuracy (%), the share of all 500 questions answere | 18.8 |
| VideoLVR Paper - TempCompass | 59.76 | Accuracy (%) | 18.2 |
| VideoLVR Paper - Video-MMMU Adaptation | 25.67 | Accuracy (%) | 18.2 |
| VideoZeroBench | 5.8 | Level-3 accuracy (%), standard end-to-end question answering | 12.5 |
| VideoLVR Paper - Video-MMMU Comprehension | 39.33 | Accuracy (%) | 9.1 |
Interactive version: theaggregate.ai/model?slug=videorft-7b · How It Works · Data refreshed daily, snapshot 2026-11-08.