RenlyH/CodeV-RL
CodeV-RL is a 7 billion parameter code-based visual agent developed by RenlyH, trained with Tool-Aware Policy Optimization (TAPO) for faithful visual reasoning. This agentic vision-language model is designed to process and reason with images by calling image operations, addressing unfaithful visual reasoning common in prior models. It achieves competitive accuracy and significantly increases faithful tool-use rates on visual search benchmarks, while also performing strongly on multimodal reasoning and math tasks.
Loading preview...
RenlyH/CodeV-RL: A Code-Based Visual Agent for Faithful Reasoning
CodeV-RL is a 7 billion parameter agentic vision-language model developed by RenlyH, specifically designed to enhance visual reasoning through code-based interactions. It utilizes Tool-Aware Policy Optimization (TAPO) to enable the model to "think with images" by dynamically calling image operations, thereby addressing the issue of unfaithful visual reasoning often found in other models. This approach allows CodeV-RL to interpret and interact with visual data more reliably and accurately.
Key Capabilities
- Faithful Visual Reasoning: Employs a code-based approach and TAPO to ensure more reliable and accurate interpretation of visual information.
- Agentic Vision-Language Model: Capable of interacting with images through programmatic operations, making it highly adaptable for visual tasks.
- Strong Performance on Visual Search: Achieves competitive accuracy and significantly improves faithful tool-use rates on visual search benchmarks.
- Multimodal Reasoning: Demonstrates robust performance across various multimodal reasoning tasks.
- Mathematical Reasoning: Shows strong capabilities in mathematical problem-solving.
Good for
- Applications requiring high-fidelity visual understanding: Ideal for scenarios where accurate and trustworthy visual interpretation is critical.
- Developing intelligent agents that interact with visual data: Its agentic nature makes it suitable for tasks involving dynamic image processing and analysis.
- Research in faithful AI and multimodal reasoning: Provides a strong baseline and methodology for advancing visual reasoning capabilities.
- Tasks involving visual search and complex visual queries: Excels in scenarios where precise tool-use and visual information extraction are paramount.