| --- |
| license: apache-2.0 |
| base_model: |
| - Qwen/Qwen2.5-VL-7B-Instruct |
| --- |
| |
| <h1 align="center">VL-Cogito</h1> |
| <p align="center"> |
| <a href="https://github.com/alibaba-damo-academy/VL-Cogito" target="_blank" rel="noopener">Website</a> |
| |
| <a href="https://huggingface.co/csyrf/VL-Cogito" target="_blank" rel="noopener"> Model </a> |
| |
| <a href="https://huggingface.co/datasets/csyrf/VL-Cogito" target="_blank" rel="noopener"> Dataset </a> |
| |
| <a href="https://arxiv.org/abs/2507.22607" target="_blank" rel="noopener">Paper</a> |
| </p> |
|
|
| The homepage of our multimodal reasoning model—VL-Cogito! |
| Inspired by the Latin word “Cogito” (“I think”), VL-Cogito is built for complex and diverse multimodal reasoning tasks, with a strong focus on autonomous thinking and adaptability. |
|
|
| **What makes VL-Cogito stand out?** |
|
|
| Progressive Curriculum Reinforcement Learning (PCuRL):Through a multi-stage, “from easy to hard” reinforcement learning approach, VL-Cogito’s reasoning abilities are significantly enhanced across a wide range of multimodal scenarios! |
|
|
| **Two key innovations:** |
| + Online difficulty weighting: Dynamically adjusts training difficulty, allowing the model to progress step by step from easier to more challenging examples. |
| + Dynamic length reward: Encourages the model to adapt the length of its reasoning process based on the complexity of each individual problem, balancing both accuracy and efficiency. |
|
|
| **Outstanding Performance:** |
|
|
| VL-Cogito demonstrates stable, state-of-the-art or superior results on mainstream multimodal reasoning benchmarks, covering mathematics, science, logic, and commonsense understanding! |
|
|
|  |
|
|
