zhangkangning/VAD_for_Qwen3.5-9b
zhangkangning/VAD_for_Qwen3.5-9b is a 9 billion parameter multimodal model based on Qwen3.5-9B, specifically enhanced for fine-grained visual question answering through Visual Attribution Distillation (VAD). This model excels at attributing responses to specific visual evidence, achieving an unweighted six-benchmark average of 79.93 on the Vision-OPD evaluation pipeline. It is designed for research in multimodal perception and on-policy distillation, offering improved visual reasoning capabilities.
Loading preview...
Overview
This model, zhangkangning/VAD_for_Qwen3.5-9b, is a 9 billion parameter multimodal large language model built upon the Qwen3.5-9B architecture. It incorporates Visual Attribution Distillation (VAD), a novel training methodology that focuses on attributing model corrections to relevant visual evidence. Unlike traditional multimodal distillation that transfers complete teacher corrections, VAD isolates and projects the teacher's correction onto a direction indicating visual support or refutation, allocating separate budgets for supported and refuted candidate tokens.
Key Capabilities & Features
- Enhanced Visual Attribution: Specifically trained to discern and respond based on fine-grained visual evidence, rather than mixing visual cues with other teacher-specific effects.
- High Performance on Visual Benchmarks: Achieves an unweighted average accuracy of 79.93 across six benchmarks (VStar, ZoomBench, HRBench-4K, HRBench-8K, MME-RealWorld-EN, MME-RealWorld-CN) under the Vision-OPD official evaluation pipeline.
- Efficient Inference: Although auxiliary teacher views are used during training, inference requires only a single image and utilizes the standard Qwen3.5-9B architecture, making it efficient for deployment.
- Robust Training: Trained using a symmetric Jensen--Shannon divergence (JSD) objective over 2 epochs with 6,241 synthetic visual question-answering examples.
Intended Use Cases
This model is primarily intended for:
- Research in Multimodal Perception: Exploring how models process and attribute information from visual inputs.
- Fine-grained Visual Question Answering: Answering complex questions that require detailed understanding and attribution of visual elements.
- On-policy Distillation Studies: Investigating advanced distillation techniques for multimodal models.
Users should be aware that performance outside reported visual benchmarks, including broad language capabilities, safety, and bias, has not been comprehensively evaluated. The model inherits the capabilities and limitations of its base, Qwen3.5-9B.