MBZUAI/Video-R2
MBZUAI/Video-R2 is a 7 billion parameter video reasoning multimodal language model (MLLM) developed by MBZUAI. Built upon Qwen2.5-VL-Instruct, it is specifically designed to produce consistent, temporally grounded, and visually faithful reasoning over dynamic video content. This model addresses common MLLM failure modes by ensuring logical consistency between reasoning and final answers and reducing over-reliance on linguistic priors, making it ideal for complex video understanding tasks.
Loading preview...
Video-R2: Consistent and Grounded Video Reasoning
Video-R2 is a 7 billion parameter multimodal language model (MLLM) from MBZUAI, specifically engineered for advanced video reasoning. It builds on the Qwen2.5-VL-Instruct base model and introduces novel post-training stages: timestamp-aware supervised fine-tuning (SFT) for explicit temporal grounding and Group Relative Policy Optimization (GRPO) to reinforce consistency and video reliance.
Key Capabilities
- Consistent Reasoning: Addresses logical inconsistencies between generated reasoning and final answers, a common issue in prior video MLLMs.
- Visually Grounded: Reduces over-reliance on linguistic priors by ensuring reasoning is strongly tied to video evidence.
- Temporal Awareness: Incorporates timestamp-aware SFT to improve explicit temporal grounding within video content.
- Diagnostic Metrics: Utilizes Think-Answer Consistency (TAC) and Video Attention Score (VAS) to measure reasoning alignment and visual reliance.
Good for
- Complex Video Understanding: Excels in scenarios requiring deep analysis of dynamic video content, such as answering intricate questions about events or actions.
- Applications Demanding High Consistency: Ideal for tasks where the model's reasoning process must logically align with its final conclusions.
- Research in Video MLLMs: Provides a strong baseline and methodology for developing more reliable and visually faithful video reasoning systems.
Video-R2 demonstrates superior performance across 11 benchmarks, including 6 reasoning-focused ones like VideoMathQA and SciVideoBench, achieving higher consistency (TAC) and video-focused reasoning (VAS) while maintaining competitive accuracy compared to models like Video-R1 and VideoChat-R1.