--- base_model: - Qwen/Qwen2-VL-2B-Instruct datasets: - tanhuajie2001/Reason-RFT-CoT-Dataset language: - en license: apache-2.0 metrics: - accuracy pipeline_tag: image-text-to-text library_name: transformers tags: - visual-reasoning --- # Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models The model was presented in the paper [Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning of Vision Language Models](https://huggingface.co/papers/2503.20752).
*The model checkpoints in our project "Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning"*

  ⭐️ Project   │   🌎 Github   │   🔥 Dataset   │   📑 ArXiv   │   💬 WeChat

  🤖 RoboBrain: Aim to Explore ReasonRFT Paradigm to Enhance RoboBrain's Embodied Reasoning Capabilities.

## ♣️ Model List | Tasks | Reason-RFT-Zero-2B | Reason-RFT-Zero-7B | Reason-RFT-2B | Reason-RFT-7B | |------------------------|---------------------------|---------------------|---------------------------|---------------------------| | Visual Counting | [🤗VC-GRPO-Zero-2B](https://huggingface.co/tanhuajie2001/Reason-RFT-Zero-Visual-Counting-Qwen2-VL-2B) | [🤗VC-GRPO-Zero-7B](https://huggingface.co/tanhuajie2001/Reason-RFT-Zero-Visual-Counting-Qwen2-VL-7B) | [🤗VC-GRPO-2B](https://huggingface.co/tanhuajie2001/Reason-RFT-Visual-Counting-Qwen2-VL-2B) | [🤗VC-GRPO-7B](https://huggingface.co/tanhuajie2001/Reason-RFT-Visual-Counting-Qwen2-VL-7B) | | Structure Perception | [🤗SP-GRPO-Zero-2B](https://huggingface.co/tanhuajie2001/Reason-RFT-Zero-Structure-Perception-Qwen2-VL-2B) | [🤗SP-GRPO-Zero-7B](https://huggingface.co/tanhuajie2001/Reason-RFT-Zero-Structure-Perception-Qwen2-VL-7B) | [🤗SP-GRPO-2B](https://huggingface.co/tanhuajie2001/Reason-RFT-Structure-Perception-Qwen2-VL-2B) | [🤗SP-GRPO-7B](https://huggingface.co/tanhuajie2001/Reason-RFT-Structure-Perception-Qwen2-VL-7B) | | Spatial Transformation | [🤗ST-GRPO-Zero-2B](https://huggingface.co/tanhuajie2001/Reason-RFT-Zero-Spatial-Transformation-Qwen2-VL-2B) | [🤗ST-GRPO-Zero-7B](https://huggingface.co/tanhuajie2001/Reason-RFT-Zero-Spatial-Transformation-Qwen2-VL-7B) | [🤗ST-GRPO-2B](https://huggingface.co/tanhuajie2001/Reason-RFT-Spatial-Transformation-Qwen2-VL-2B) | [🤗ST-GRPO-7B](https://huggingface.co/tanhuajie2001/Reason-RFT-Spatial-Transformation-Qwen2-VL-7B) | | ***Embodied Tasks*** | 🤖 *Stay Turned* | 🤖 *Stay Turned* | 🤖 *Stay Turned* | 🤖 *Stay Turned* | ## 🔥 Overview Visual reasoning abilities play a crucial role in understanding complex multimodal data, advancing both domain-specific applications and artificial general intelligence (AGI). Existing methods improve VLM reasoning via Chain-of-Thought (CoT) supervised fine-tuning, using meticulously annotated training data to enhance visual reasoning capabilities. However, this training paradigm may lead to overfitting and cognitive rigidity, restricting the model's ability to transfer visual reasoning skills across domains and limiting its real-world applicability. To address these limitations, we propose **Reason-RFT**, a novel reinforcement fine-tuning framework that significantly enhances generalization capabilities in visual reasoning tasks. **Reason-RFT** introduces a two-phase training framework for visual reasoning: (1) Supervised Fine-Tuning (SFT) with curated Chain-of-Thought (CoT) data activates the reasoning potential of Vision-Language Models (VLMs), followed by (2) Group Relative Policy Optimization (GRPO)-based reinforcement learning that generates multiple reasoning-response pairs, significantly enhancing generalization in visual reasoning tasks. To evaluate **Reason-RFT**'s visual reasoning capabilities, we reconstructed a comprehensive dataset spanning visual counting, structure perception, and spatial transformation, serving as a benchmark to systematically assess visual cognition, geometric understanding, and spatial generalization. Experimental results demonstrate Reasoning-RFT's three key advantages: **(1) Performance Enhancement**: achieving state-of-the-art results across multiple tasks, outperforming most mainstream open-source and proprietary models; **(2) Generalization Superiority**: consistently maintaining robust performance across diverse tasks and domains, outperforming alternative training paradigms; **(3) Data Efficiency**: excelling in few-shot learning scenarios and surpassing full-dataset SFT baselines; **Reason-RFT** introduces a novel paradigm in visual reasoning, significantly advancing multimodal research.
## 🗞️ News - **`2025-09-18`**: 🔥🔥🔥 **Reason-RFT** gets accepted to NeurIPS 2025! See you in Mexico City and San Diego, USA! - **`2025-06-06`**: 🤖 We're excited to announce the release of our more powerful [RoboBrain 2.0](https://github.com/FlagOpen/RoboBrain2.0) using Reason-RFT. - **`2025-04-13`**: ✨ We released our [model zoo](https://github.com/tanhuajie/Reason-RFT?tab=readme-ov-file#--model-zoo) to huggingface. - **`2025-04-12`**: ⭐️ We released our [Models](https://huggingface.co/tanhuajie2001/Reason-RFT-Spatial-Transformation-Qwen2-VL-2B) to huggingface for [General Visual Reasoning Tasks](#GeneralVisualTasks). - **`2025-04-04`**: 🤗 We released our [datasets](https://huggingface.co/datasets/tanhuajie2001/Reason-RFT-CoT-Dataset/) to huggingface for [General Visual Reasoning Tasks](#GeneralVisualTasks). - **`2025-04-02`**: 🔥 We released codes and scripts for training/evaluation on [General Visual Reasoning Tasks](#GeneralVisualTasks). - **`2025-03-29`**: 🌍 We released the [repository](https://github.com/tanhuajie/Reason-RFT/) and [roadmap](#RoadMap) for **Reason-RFT**. - **`2025-03-26`**: 📑 We released our initial [ArXiv paper](https://arxiv.org/abs/2503.20752/) of **Reason-RFT**. ## ⭐️ Sample Usage The following code snippet demonstrates how to perform quick inference with the model. For more details, you could refer to [Github](https://github.com/tanhuajie/Reason-RFT). ```python # git clone https://github.com/tanhuajie/Reason-RFT import numpy as np import torch from longvu.builder import load_pretrained_model from longvu.constants import ( DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_INDEX, ) from longvu.conversation import conv_templates, SeparatorStyle from longvu.mm_datautils import ( KeywordsStoppingCriteria, process_images, tokenizer_image_token, ) from decord import cpu, VideoReader # This is an example, replace with the correct path to your downloaded checkpoint tokenizer, model, image_processor, context_len = load_pretrained_model( "./checkpoints/longvu_qwen", None, "cambrian_qwen", ) model.eval() video_path = "./examples/video1.mp4" # Replace with your image/video path qs = "Describe this video in detail" # Replace with your query # For image input, replace this section with image loading and processing vr = VideoReader(video_path, ctx=cpu(0), num_threads=1) fps = float(vr.get_avg_fps()) frame_indices = np.array([i for i in range(0, len(vr), round(fps),)]) video = [] for frame_index in frame_indices: img = vr[frame_index].asnumpy() video.append(img) video = np.stack(video) image_sizes = [video[0].shape[:2]] video = process_images(video, image_processor, model.config) video = [item.unsqueeze(0) for item in video] qs = DEFAULT_IMAGE_TOKEN + " " + qs conv = conv_templates["qwen"].copy() conv.append_message(conv.roles[0], qs) conv.append_message(conv.roles[1], None) prompt = conv.get_prompt() input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).to(model.device) stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2 keywords = [stop_str] stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, input_ids) with torch.inference_mode(): output_ids = model.generate( input_ids, images=video, image_sizes=image_sizes, do_sample=False, temperature=0.2, max_new_tokens=128, use_cache=True, stopping_criteria=[stopping_criteria], ) pred = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip() print(f'Generated response: {pred}') ``` ## 📑 Citation If you find this project useful, welcome to cite us. ```bib @article{tan2025reason, title={Reason-RFT: Reinforcement Fine-Tuning for Visual Reasoning}, author={Tan, Huajie and Ji, Yuheng and Hao, Xiaoshuai and Lin, Minglan and Wang, Pengwei and Wang, Zhongyuan and Zhang, Shanghang}, journal={arXiv preprint arXiv:2503.20752}, year={2025} } @article{team2025robobrain, title={Robobrain 2.0 technical report}, author={Team, BAAI RoboBrain and Cao, Mingyu and Tan, Huajie and Ji, Yuheng and Lin, Minglan and Li, Zhiyu and Cao, Zhou and Wang, Pengwei and Zhou, Enshen and Han, Yi and others}, journal={arXiv preprint arXiv:2507.02029}, year={2025} } @article{ji2025robobrain, title={RoboBrain: A Unified Brain Model for Robotic Manipulation from Abstract to Concrete}, author={Ji, Yuheng and Tan, Huajie and Shi, Jiayu and Hao, Xiaoshuai and Zhang, Yuan and Zhang, Hengyuan and Wang, Pengwei and Zhao, Mengdi and Mu, Yao and An, Pengju and others}, journal={arXiv preprint arXiv:2502.21257}, year={2025} } ```