| license: apache-2.0 | |
| library_name: transformers | |
| pipeline_tag: video-text-to-text | |
| tags: | |
| - multimodal | |
| - agent | |
| - reinforcement-learning | |
| # PyVision-Video-7B-RL | |
| [**PyVision-RL: Forging Open Agentic Vision Models via RL**](https://huggingface.co/papers/2602.20739) | |
| PyVision-Video-7B-RL is an open-weight agentic multimodal model post-trained from [Qwen2.5-VL-7B](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) using reinforcement learning. | |
| - **Project Page:** [agent-x.space/pyvision-rl](https://agent-x.space/pyvision-rl/) | |
| - **GitHub Repository:** [agents-x-project/PyVision-RL](https://github.com/agents-x-project/PyVision-RL) | |
| - **Paper:** [arXiv:2602.20739](https://arxiv.org/abs/2602.20739) | |
| ## Overview | |
| Reinforcement learning for agentic multimodal models often suffers from interaction collapse, where models learn to reduce tool usage and multi-turn reasoning. **PyVision-RL** is a framework designed to stabilize training and sustain interaction by combining an oversampling-filtering-ranking rollout strategy with an accumulative tool reward. | |
| **PyVision-Video** specifically addresses the challenge of video reasoning using **on-demand context construction**. It selectively samples task-relevant frames during the reasoning process to significantly reduce visual token usage while maintaining high performance on complex multimodal agentic tasks. | |
| ## Citation | |
| ```bibtex | |
| @article{zhao2026pyvisionrl, | |
| title={PyVision-RL: Forging Open Agentic Vision Models via RL.}, | |
| author={Zhao, Shitian and Lin, Shaoheng and Li, Ming and Zhang, Haoquan and Peng, Wenshuo and Zhang, Kaipeng and Wei, Chen}, | |
| journal={arxiv preprint arxiv:2602.20739}, | |
| year={2026}, | |
| } | |
| ``` |