SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI developer platformsVision-R1

Vision-R1

Vision-R1 is a project for enhancing the reasoning capabilities of multimodal large language models (MLLMs), proposed and open-sourced by the relevant research team. It is the first work to explore how to effectively apply R1-like reinforcement learning to multimodal large language models, incentivizing the model to produce reasoning capabilities through a combination of cold-start initialization and reinforcement learning training. The project provides the cold-start dataset Vision-R1-cold, the reinforcement learning dataset Vision-R1-rl, and model weights of multiple scales (Vision-R1-CI-7B, Vision-R1-7B, Vision-R1-32B, Vision-R1-72B), and has open-sourced the inference code, evaluation code, and training scripts. This work has been accepted by ICLR 2026.

Open website

About this tool

Vision-R1 is the official code repository for the paper "Vision-R1: Incentivizing Reasoning Capability in Multimodal Large Language Models," and has been accepted by ICLR 2026. The project is dedicated to exploring how to effectively apply a reinforcement learning paradigm similar to DeepSeek-R1 to multimodal large language models (MLLMs), thereby incentivizing their reasoning capabilities.

[Purpose] The core goal of Vision-R1 is to improve the reasoning capabilities of multimodal large language models. Through a two-stage process combining cold-start initialization and reinforcement learning (RL) training, it builds a multimodal model with reasoning capabilities.

[Main Capabilities]

  • Cold-start initialization: First, use existing MLLMs and DeepSeek-R1 to generate a high-quality multimodal chain-of-thought (CoT) dataset as cold-start initialization data for the base MLLM, yielding Vision-R1-CI.
  • Reinforcement learning training: Perform RL training on the basis of Vision-R1-CI to obtain the final reasoning model Vision-R1. The project proposes the PTST strategy, gradually relaxing the context length limit (4K, 8K, 16K tokens) to acquire more complex reasoning processes in a progressive manner.
  • Interleaved reasoning generation: Supports interleaved reasoning generation and display.
  • Model scales: Provides weights such as Vision-R1-CI-7B, Vision-R1-7B, Vision-R1-32B, and Vision-R1-72B. Among them, 32B and 72B used additional data during RL training.
  • Datasets: Provides the cold-start dataset Vision-R1-cold and the reinforcement learning dataset Vision-R1-rl. The images of Vision-R1-cold can come from LLaVA-CoT-100k and Mulberry-SFT, and need to be placed in the llava_cot_images and mulberry_images directories respectively. Vision-R1-rl uses the geometry3k format.
  • Inference and evaluation: Open-sourced the inference code, scripts, and evaluation code, supports vLLM deployment (requires vllm>0.7.2), can launch an OpenAI-compatible API service, and also supports local inference. During inference, useSame temperature and top_p as DeepSeek-R1.
  • Training tools: In the cold-start phase, use LLaMA-Factory to train Vision-R1-CI, with the training script located in train/cold_start; in the RL phase, training can be completed using EasyR1 or verl.

[Usage]

  1. Download the cold-start dataset Vision-R1-cold, and obtain images from LLaVA-CoT-100k and Mulberry-SFT, placing them in the corresponding directories.
  2. Use LLaMA-Factory and the script under train/cold_start to train the cold-start model; 8*4 or 8 GPUs with 80G VRAM can be used. The dataset path in train.yaml needs to be replaced.
  3. Use the Vision-R1-rl dataset to perform RL training via EasyR1 or verl.
  4. For inference, you can use the scripts/inference.sh script provided by the project, or use vLLM to deploy an OpenAI-compatible API service and make conversational calls.
  5. The evaluation code is located in the evaluation/ directory.

[Applicable Scenarios] Suitable for research on the reasoning capabilities of multimodal large language models, tasks requiring multimodal chain-of-thought such as mathematical reasoning, and scientific research and engineering scenarios where users wish to reproduce or conduct secondary development based on Vision-R1.

[Limitations and Notes] The paper points out that directly applying RL to MLLMs makes it difficult to effectively incentivize strong reasoning capabilities; Vision-R1-Zero without initialization training struggles to generalize on limited data; Vision-R1-CI has an overthinking optimization problem, preferring shorter CoTs, and in subsequent RL training the reasoning steps become longer but performance decreases. Vision-R1 shortens the CoT in the early stage of RL training to refine the correct thinking process, and PTST enables the model to gradually acquire more complex reasoning processes. In addition, the project mentions that it is trying to extend training to low-resource settings, with the goal of completing the entire cold-start and RL training process with 8 GPUs. The training and release progress of the 72B model, the complete dataset and code, the weight release status, etc. are subject to the actual information in the repository; some content and materials are marked as planned or in progress.

Related tools

Browse category