Video Question Answering on VideoEspresso (test)
0.278AccuracyREVISE (Reinforced Fine-Tuning)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| REVISE (Reinforced Fine-Tuning)Backbone=Qwen-2.5-VL-3B, Reasoning Category=Reinforced Fine-Tuning (RFT)2026.02 | 0.278 | 4.1 | 1.37 | 1.02 | |
| Supervised Format Fine-TuningReasoning Category=supervised format fine-tuning2026.02 | 0.213 | 5 | 1.52 | 1.67 | |
| Plug-and-PlayReasoning Category=plug-and-play without updates2026.02 | 0.201 | 5.2 | 1.86 | 1.73 | |
| Direct ReasoningReasoning Category=direct video-understanding2026.02 | 0.126 | 8 | 1 | 1.02 |