HaochenWang/TreeVGR-7B
TreeVGR-7B is a 7 billion parameter visual grounded reasoning model developed by Haochen Wang and collaborators, initialized from Qwen2.5-VL-7B. It is specifically designed for "thinking with images" capabilities, focusing on traceable visual evidence and second-order reasoning. The model excels at tasks requiring precise localization and understanding of object interactions and spatial hierarchies, achieving state-of-the-art performance on the TreeBench benchmark with a 32768 token context length.
Loading preview...
TreeVGR-7B: Traceable Visual Grounded Reasoning
TreeVGR-7B is a 7 billion parameter visual grounded reasoning model, initialized from Qwen2.5-VL-7B, developed by Haochen Wang and collaborators. It is specifically engineered to enhance "thinking with images" capabilities by focusing on traceable visual evidence and second-order reasoning. This model addresses a critical gap in evaluating visual reasoning by introducing TreeBench, a diagnostic benchmark designed for complex scenes, subtle target perception, and traceable evidence via bounding box evaluation.
Key Capabilities
- Traceable Visual Evidence: Supervises localization and reasoning jointly using reinforcement learning, enabling accurate localizations and explainable reasoning pathways.
- Second-Order Reasoning: Excels at understanding object interactions and spatial hierarchies, moving beyond simple object localization.
- High-Quality Visual Perception: Designed to handle complex scenes and subtle targets, as evaluated by the TreeBench benchmark.
- Improved Benchmarks: Demonstrates significant improvements on V* Bench (+16.8), MME-RealWorld (+12.6), and TreeBench (+13.4) compared to its base model.
Good for
- Advanced Visual Question Answering: Ideal for tasks requiring deep understanding of visual content, including object relationships and spatial reasoning.
- Research in Visual Grounded Reasoning: Provides a strong baseline and methodology for further advancements in models that can "think with images" with explainable outputs.
- Applications Requiring Traceable AI: Useful in scenarios where not just the answer, but also the visual evidence supporting it, is crucial.