SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsOther AI toolsRevisual-R1

Revisual-R1

Revisual-R1 is a 7B open-source multimodal language model, and its related paper has been accepted by ICLR 2026. It adopts a three-stage curriculum training approach: cold-start pretraining, multimodal reinforcement learning, and pure-text reinforcement learning, with the goal of achieving faithful, concise, and self-reflective state-of-the-art performance in visual and textual reasoning. The model achieves SOTA performance among open-source 7B models on multiple challenging visual math and reasoning benchmarks, while retaining strong general text capabilities. The models (Coldstart and final versions) have been open-sourced on Hugging Face, and the paper is available on arXiv. The project team also plans to open-source the GRAMMAR dataset containing high-quality cold-start data, and to advance the training and release of 32B and 3B versions. The repository provides dependency installation commands, as well as methods for merging Hugging Face-format checkpoints, direct evaluation through the OpenAI API, and deployment and evaluation of local models using lmdeploy.

Open website

About this tool

Purpose

Revisual-R1 is a 7B-parameter open-source multimodal language model (MLLM), designed to address the common challenges in cultivating complex multimodal reasoning capabilities. The related work has been accepted by ICLR 2026, focusing on exploring how to unlock robust, self-reflective multimodal reasoning through data strategies and algorithm optimization, rather than simply scaling up model size.

Main Capabilities

  • It adopts a three-stage curriculum training approach: a strategic high-difficulty pure-text cold-start stage, a multimodal reinforcement learning stage, and a pure-text reinforcement learning refinement stage.
  • The multimodal reinforcement learning stage uses GRPO and stabilizes training with the proposed Prioritized Advantage Distillation (PAD) mechanism, together with rule-based rewards (including Efficient-Length Reward).
  • It achieves SOTA among open-source 7B models on multiple challenging visual math and reasoning benchmarks, while retaining strong general text capabilities.
  • The model can generate longer responses, continuously hypothesizing, reflecting, verifying, and correcting, and finally giving an answer, while also providing a summary-style response.

Usage

The repository provides the following installation and usage paths (specific commands and parameters are subject to the repository files):

  • Install dependencies; if problems are encountered when installing flash-attn, the repository recommends referring to flash-attn for installation; if problems are encountered when connecting to Hugging Face, you can use export HF_ENDPOINT= https://hf-mirror.com.
  • Merge Hugging Face-format checkpoints.
  • Directly evaluate the model through the OpenAI API.
  • Use lmdeploy to deploy and evaluate a local model; you can also modify the parameters in inference/inference.sh .

Applicable Scenarios

It is suitable for visual math reasoning, multimodal reasoning research, and scenarios that require the model to balance performance across visual and text tasks and possess self-reflection and multi-hop reasoning capabilities.

Limitations and Not Yet Provided

  • The repository does not provide information related to fees, pricing, or commercial licensing.
  • The training dataset (Coldstart) has not yet been released, and the project team says it will be completed in the coming weeks; the GRAMMAR dataset is also planned to be open-sourced within two weeks.
  • The 32B and 3B Revisual-R1 are still in training and will be made public after release.
  • Specific hardware requirements, complete installation commands, evaluation parameters, and deployment details are subject to the repository files, and this introduction does not elaborate on them or does not yet provide them.

Citation and Contact

If the work is helpful to research, it can be cited according to the citation information given in the repository. Questions, feedback, or collaboration can be sent by email to csfufu0728@gmail.com.

Related tools

Browse category