Step-R1-V-Mini - Step-R1's latest multimodal inference model

What is Step-R1-V-Mini?

Step-R1-V-Mini is StepStar's latest multimodal reasoning model. It supports image-text input and text output, possesses excellent instruction compliance and versatility, and can accurately perceive images to complete complex reasoning tasks. The model excels in visual reasoning and is also among the top performers in mathematical, code, and text reasoning. It employs multimodal joint reinforcement learning, introducing a verifiable reward mechanism in the image space based on the PPO strategy to improve generalization and robustness. Training with multimodal synthetic data effectively solves the seesaw problem in training.

Main functions of Step-R1-V-Mini

• Multimodal input and outputIt supports image and text input and text output, can process image and text information, outputs reasoning results in text form, and has good command compliance and versatility.
• High-precision image perception and reasoningIt can perceive images with high precision and complete complex reasoning tasks, such as identifying specific locations through images, analyzing food images and generating detailed recipes. It ranks first in China on the MathVision visual reasoning leaderboard.
• Solving mathematical problemsIt can construct a reasonable reasoning chain and plan and solve complex mathematical problems step by step, including Olympiad math problems and geometry problems.
• Logical reasoning analysisTry different problem-solving approaches independently, ask yourself questions to ensure that all good solutions are enumerated, and check for any omissions before submitting the paper.
• Solutions to complex algorithm problemsAble to correctly solve algorithm problems rated "Hard" on the LeetCode platform.
• Code Logic ConstructionStep by step, analyze user needs and intentions, construct code logic, and interweave analysis and verification of the current code snippet during code writing.
• Literary creation: Deeply understand users' expressive needs, analyze the requirements of creative themes and literary subjects, endow things with symbolic meaning on the level of human emotions, and increase personalized and innovative expressive styles.

The technical principles of Step-R1-V-Mini

• Multimodal joint reinforcement learningStep-R1-V-Mini is a reinforcement learning method based on the Proximal Policy Optimization (PPO) policy. PPO is an on-policy algorithm that updates the model in real time by generating samples online. In the image space, the model introduces a verifiable reward mechanism to address the complex inference links in image space, which are prone to causing confusion in correlation and causal inference errors. Compared to traditional methods such as Direct Preference Optimization (DPO), PPO is more generalizable and robust when dealing with complex links in the image space.
• High-quality multimodal data synthesisSince feedback signals for multimodal data are relatively difficult to obtain, Step-R1-V-Mini designs a large number of multimodal data synthesis links based on environmental feedback. By synthesizing scalable multimodal inference data and combining it with PPO-based reinforcement learning training, the model can simultaneously improve its text and visual inference capabilities. This effectively avoids the "seesaw" problem in training, where different modal capabilities rise and fall at opposite ends.
• Cold start and multi-stage reinforcement learningThe Step-R1-V-Mini training process also involves cold start and multi-stage reinforcement learning. First, the base model is fine-tuned using high-quality cold start data (such as CoT data) to obtain an initial model. Then, large-scale reinforcement learning training is performed on the model to enable it to exhibit inference capabilities. Next, high-quality SFT data is generated using the trained model, mixed with data from other domains, and SFT training is performed again. Finally, the final reinforcement learning is performed using data from all domains to obtain the final model.

How to use Step-R1-V-Mini

• Access Step AI Web PageStep-R1-V-Mini is now officially available on Step AI. You can visit the official website to directly select the model for inference tasks.
• Calling API InterfaceFor developers or enterprise users, Step-R1-V-Mini provides API interfaces on the Step-R1 Starry Sky Open Platform. Detailed API documentation and calling methods can be obtained by accessing the Step-R1 Starry Sky Open Platform.
• Visual reasoningUsers can upload images and enter related questions. The model can perceive the images with high accuracy and complete complex reasoning tasks, such as identifying locations in the images, calculating the number of objects, and recognizing recipes.
• Mathematical and Logical ReasoningInput a mathematical problem or a logical reasoning question, and the model can construct a reasonable reasoning chain and solve it step by step.

Application scenarios of Step-R1-V-Mini

• Image recognition and analysisIt can perceive images with high precision and complete complex reasoning tasks. The model can quickly identify elements in the image, combine information such as color and objects, and comprehensively determine that the location is Wembley Stadium, providing possible information about the two sides in the match.
• Calculation of the number of objectsInput a diagram containing objects of different shapes, colors, and positions. The model can identify each object and perform logical reasoning to determine the number of remaining objects.
• Recipe recognitionInput a picture of food, and the model can accurately identify the dish and dipping sauce, and list the specific amounts required.
• Multimodal data fusionIt can process multimodal data containing text and images and generate comprehensive reasoning results.
• Cross-modal reasoningTransforming images into formal textual descriptions enables language models to accurately process and reason about images.