V-Zero

V-Zero introduces a novel approach to fine-grained visual reasoning that eliminates the need for annotated answer labels by utilizing a contrastive evidence-gated distillation method. The framework involves a student model that samples reasoning trajectories from the full image, while a teacher model evaluates these trajectories by comparing positive and negative visual evidence views. This contrastive analysis helps determine the degree to which a trajectory is grounded in actual visual evidence, allowing for selective distillation of teacher signals that are more likely to be supported by relevant visual data. By avoiding reliance on language priors and focusing on evidence-based reasoning, V-Zero provides dense token-level supervision during training without altering the standard full-image inference process. The approach leverages on-policy rollouts and evidence-aware distillation to refine model performance in identifying small objects, reading local text, and comparing subtle visual attributes, thus addressing challenges in multimodal large language models without the need for human-written ground-truth answers.