智源 FlagOpen 开源 Reason-RFT：面向视觉推理的强化微调框架

FlagOpen; Reason-RFT

Repository navigation

NotificationsYou must be signed in to change notification settings; Fork2 (2); Star4 (4)You must be signed in to star a repository

Repository files navigation

README; Apache-2.0 license

⭐️ Project │ 🌎 Dataset │ 🤗 Models │ 📑 Paper │ 💬 WeChat

🤖 RoboBrain 2.0: Explore Embodied Visual Reasoning Capabilities with Reason-RFT.

🔥 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 while surpassing full-dataset SFT baselines; Reason-RFT introduces a novel paradigm in visual reasoning, significantly advancing multimodal research.

🎯 RoadMap

Support different VLMs: RoboBrain, Qwen2-VL series, Llava-VL series.

Explore an efficient training paradigm to enhance RoboBrain's embodied reasoning capabilities; Support General Visual Reasoning Tasks:

Data download and preparation: Please refer to General Visual Reasoning Tasks; Training and evaluating for Visual Counting: Please refer to Visual Counting Section; Training and evaluating for Struction Perception: Please refer to Struction Perception Section; Training and evaluating for Spatial Transformation: Please refer to Spatial Transformation Section; Support Embodied Visual Reasoning Tasks:

Data generation and preparation: Please refer to Embodied Visual Reasoning Tasks; Training and evaluating for Embodied Planning: Please refer to Embodied Planning Section; Training and evaluating for Embodied Affordance: Please refer to Embodied Affordance Section; Training and evaluating for Embodied Trajectory: Please refer to Embodied Trajectory Section; Training and evaluating for Embodied Pointing: Please refer to Embodied Pointing Section

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 using Reason-RFT; 2025-04-13: ✨ We released our model zoo to huggingface; 2025-04-04: 🤗 We released our datasets to huggingface for General Visual Reasoning Tasks; 2025-04-02: 🔥 We released codes and scripts for training/evaluation on General Visual Reasoning Tasks; 2025-03-29: 🌍 We released the repository and roadmap for Reason-RFT; 2025-03-26: 📑 We released our initial ArXiv paper of Reason-RFT

⭐️ Pipeline

🤗 Model Zoo

🎲 General Visual Reasoning Tasks

Then, your local directory should be like:

Note: If you want to implement training process for our three main tasks, only three Meta-Json files -- 'A1-Spatial-Transformation-train-60k-cot.json', 'A2-Structure-Perception-train-4k5-cot.json' and 'A3-Visual-Counting-train-35k-cot.json' -- are needed respectively. Then, all the training-used json files can be constructed by these three Meta-Json files above. At this step, you should write the simple scripts to complete the construction for your own training-used json files directly, according to the sample format as shown below.

📑 1. For ANS-SFT training, we use ShareGPT format to refactor each sample:

📑 2. For COT-SFT training, we also use ShareGPT format to refactor each sample:

📑 3. For Reason-RFT-Zero training, we use RL format below to refactor each sample:

📑 4. For Reason-RFT training, we use COT-SFT format to refactor 1.6k samples for STAGE-1, and use RL format to refactor the rest samples or full samples for STAGE-2.

Specifically, in STAGE 2, we refactor full samples for Structure-Perception task due to its too limited training samples (only 4.5k), while we refactor only the rest samples for Spatial-Transformation task and Visual-Counting task.

📚 Training

Note: Please change the dataset, pre-trained model and image path in the scripts above.

🔭 Evaluation

Note: Please change the checkpoint path in the scripts above.

⭐️ Results in our paper

1. Results of Spatial Transformation Task:

2. Full Results of Three Tasks:

Note: For more detail results, please refer to our main paper and appendix

🤖 Embodied Visual Reasoning Tasks

We apply Reason-RFT to train more powerful RoboBrain 2.0. Please refer to RoboBrain 2.0 Github for more details. Here is a simplified comparison result (not the final version):

📑 Citation

If you find this project useful, welcome to cite us.