Senqiao/VisionThink-Efficient
Senqiao/VisionThink-Efficient is a 7 billion parameter vision-language model (VLM) developed by Senqiao Yang et al. with a 32768 token context length. It is trained via reinforcement learning to autonomously reduce visual tokens, achieving enhanced performance and efficiency on general Visual Question Answering (VQA) tasks. This model excels particularly in fine-grained benchmarks, including OCR-related tasks, by maintaining performance while reducing visual tokens by 50%.
Loading preview...
VisionThink-Efficient: Smart and Efficient VLM
Senqiao/VisionThink-Efficient is a 7 billion parameter vision-language model (VLM) developed by Senqiao Yang and collaborators. This model distinguishes itself by leveraging reinforcement learning to autonomously determine when to reduce visual tokens, leading to significant improvements in efficiency without compromising performance.
Key Capabilities & Features
- Autonomous Visual Token Reduction: Utilizes reinforcement learning to intelligently reduce visual tokens, optimizing for efficiency.
- Enhanced Fine-Grained VQA: Achieves notable performance gains on fine-grained benchmarks, including those involving Optical Character Recognition (OCR).
- Performance with Efficiency: Improves performance on general VQA tasks while reducing visual tokens by 50%, matching 102% of the original model's performance across nine benchmarks.
- Efficient Reasoning: Inspires further research into efficient reasoning VLMs by demonstrating strong performance and efficiency through simple input image resizing.
Good For
- Applications requiring efficient Visual Question Answering where computational resources or inference speed are critical.
- Tasks involving fine-grained visual understanding, such as OCR or detailed object recognition.
- Researchers and developers interested in efficient VLM architectures and reinforcement learning applications in vision-language tasks.