CodeGoat24/WorldReward-qwen35-9b
WorldReward-qwen35-9b is a 9 billion parameter reward model developed by CodeGoat24, designed for camera-conditioned world models. It excels at evaluating video quality based on action, appearance, and motion, achieving 77.63% agreement on action, 81.32% on appearance, and 73.03% on motion on the WorldReward-Bench dataset. This model is specifically optimized for assessing the quality of generated videos in simulated environments, offering a robust metric for world model development.
Loading preview...
WorldReward-qwen35-9b: Reward Modeling for Camera-Conditioned World Models
WorldReward-qwen35-9b is a 9 billion parameter reward model developed by CodeGoat24, specifically designed to evaluate the quality of videos generated by camera-conditioned world models. This model assesses video quality across three key dimensions: Action, Appearance, and Motion.
Key Capabilities
- Comprehensive Video Evaluation: Provides a unified metric for judging video realism and consistency in simulated environments.
- High Agreement with Human Labels: Achieves strong agreement with human preferences on the WorldReward-Bench dataset, with scores of 77.63% for Action, 81.32% for Appearance, and 73.03% for Motion.
- Superior Performance: Outperforms other models like GPT-5.5, Gemini-3.1-Pro, and various Qwen3.5-VL variants in evaluating video quality, particularly in action and motion assessment.
- Camera-Conditioned: Tailored for scenarios where camera parameters influence the generated world, making it ideal for robotics and embodied AI research.
Good For
- Developing and Benchmarking World Models: Provides a robust reward signal for training and evaluating generative world models.
- Reinforcement Learning: Can be integrated into RL pipelines to guide agents in generating more realistic and coherent video sequences.
- Video Quality Assessment: Offers an automated method for comparing the output of different video generation systems in specific contexts.