🎥 $\text{VideoRFT}$: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning

  📖 ArXiv    │   📀 CoT Dataset    │   📀 RL Dataset    │   🤗 Models

📰 News

• [2025/09/19] Our paper has been accepted to NeurIPS 2025 🎉!
• [2025/06/01] We released our 3B Models (🤗VideoRFT-SFT-3B and 🤗VideoRFT-3B) to huggingface.
• [2025/05/25] We released our 7B Models (🤗VideoRFT-SFT-7B and 🤗VideoRFT-7B) to huggingface.
• [2025/05/20] We released our Datasets (📀CoT Dataset and 📀RL Dataset) to huggingface.
• [2025/05/18] Our paper is released on ArXiv, and we have open-sourced our code on GitHub!

🔎 Overview

Reinforcement fine-tuning (RFT) has shown great promise in achieving humanlevel reasoning capabilities of Large Language Models (LLMs), and has recently been extended to MLLMs. Nevertheless, reasoning about videos, which is a fundamental aspect of human intelligence, remains a persistent challenge due to the complex logic, temporal and causal structures inherent in video data. To fill this gap, we propose $\textbf{VideoRFT}$, a novel approach that extends the RFT paradigm to cultivate human-like video reasoning capabilities in MLLMs. $\textbf{VideoRFT}$ follows the standard two-stage scheme in RFT: supervised fine-tuning (SFT) with chain-of-thought (CoT) annotations, followed by reinforcement learning (RL) to improve generalization. A central challenge to achieve this in the video domain lies in the scarcity of large-scale, high-quality video CoT datasets. We address this by building a fully automatic CoT curation pipeline. First, we devise a cognitioninspired prompting strategy to elicit a reasoning LLM to generate preliminary CoTs based solely on rich, structured, and literal representations of video content. Subsequently, these CoTs are revised by a visual-language model conditioned on the actual video, ensuring visual consistency and reducing visual hallucinations. This pipeline results in two new datasets $-$ VideoRFT-CoT-102K for SFT and VideoRFT-RL-310K for RL. To further strength the RL phase, we introduce a novel semantic-consistency reward that explicitly promotes the alignment between textual reasoning with visual evidence. This reward encourages the model to produce coherent, context-aware reasoning outputs grounded in visual input. Extensive experiments show that $\textbf{VideoRFT}$ achieves state-of-the-art performance on six video reasoning benchmarks.

✨ Methodology

To overcome the scarcity of video CoTs, we develop a scalable, cognitively inspired pipeline for high-quality video CoT dataset construction.

To further strength the RL phase, we introduce a novel semantic-consistency reward that explicitly promotes the alignment between textual reasoning with visual evidence.

📀 Datasets

Based on above pipeline, we construct two large-scale datasets, i.e., 📀VideoRFT-CoT-102K and 📀VideoRFT-RL-310K.

🛠️ Set up

Requirements

• Python >= 3.11
• Pytorch >= 2.5.1
• transformers == 4.51.3
• vLLM == 0.7.3
• trl == 0.16.0

Installation

git clone https://github.com/QiWang98/VideoRFT
cd VideoRFT

# Create and activate environment
conda create -n VideoRFT python=3.11 
conda activate VideoRFT
bash setup.sh

# Install decord for improved video processing
cd src/qwen-vl-utils
pip install -e .[decord]

🚀 Training

Supervised Fine-Tuning (SFT)

We begin with supervised fine-tuning on the VideoRFT-CoT dataset for one epoch:

bash ./src/scripts/run_sft_video.sh

This step can be skipped by directly using our pretrained SFT models, available at 🤗VideoRFT-SFT-7B or 🤗VideoRFT-SFT-3B.

Reinforcement Learning (RL)

Next, perform reinforcement learning using the VideoRFT-RL dataset:

bash ./src/scripts/run_grpo_video.sh

To enable faster training via vLLM acceleration:

bash ./src/scripts/run_grpo_vllm_qwen25vl.sh
``$

> **\text{Note}:** \text{During} \text{training}, \text{we} \text{adopt} \text{the} \text{following} \text{settings} \text{for} \text{efficiency}:

* **\text{VIDEO} \text{PIXELS}**: 128  \times  28  \times  28
* **\text{FPS} \text{FRAMES}**: 16

\text{All} \text{frame}-\text{related} \text{configurations} \text{can} \text{be} \text{adjusted} \text{in} $src/qwen-vl-utils`.

## 📈 Inference & Evaluation

> During inference, we increase the maximum frame resolution and length to boost performance:

* **VIDEO PIXELS**: 256 × 28 × 28
* **FPS FRAMES**: 32

You can configure these parameters in `src/qwen-vl-utils`.

> We evaluate all models under a unified decoding configuration following the official Qwen2.5-VL demo:

* `top_p = 0.001`
* `temperature = 0.01`

### Evaluation Procedure

1. Download preprocessed evaluation JSONs from: \[[🤗 eval](https://huggingface.co/datasets/Video-R1/Video-R1-eval)]

2. Download the video data from the official sites of each benchmark and organize them as specified in the JSON files.

3. Run the evaluation across all benchmarks:

```bash
bash ./src/eval_bench.sh

🙏 Acknowledgements

We gratefully acknowledge the contributions of the open-source community, particularly DeepSeek-R1, Open-R1, and R1-V.

📚 Citations

If you find this work helpful, please consider citing:

@article{VideoRFT,
  title={VideoRFT: Incentivizing Video Reasoning Capability in MLLMs via Reinforced Fine-Tuning},
  author={Wang, Qi and Yu, Yanrui and Yuan, Ye and Mao, Rui and Zhou, Tianfei},
  booktitle={NeurIPS},
  year={2025}
}