Vision-Zero: Scalable VLM Self-Improvement via Strategic Gamified Self-Play

Researchers introduce Vision-Zero, a self-improving AI framework that trains vision-language models through competitive games without requiring human-labeled data. The system uses strategic self-play and can work with arbitrary images, achieving state-of-the-art performance on reasoning and visual understanding tasks while reducing training costs.

• →Vision-Zero eliminates the need for costly human verification and manually curated datasets in vision-language model training.
• →The framework uses competitive 'Who Is the Spy' style games to enable models to generate their own training data autonomously.
• →The system works with arbitrary images including synthetic scenes, charts, and real-world images, showing strong generalization capabilities.
• →Iterative Self-Play Policy Optimization prevents performance plateaus common in self-play training methods.
• →Despite using no labeled data, Vision-Zero achieves state-of-the-art performance surpassing annotation-based methods.